(combine_givs): When combining a DEST_REG giv with its only use,
[gcc.git] / gcc / loop.c
1 /* Perform various loop optimizations, including strength reduction.
2 Copyright (C) 1987, 88, 89, 91-5, 1996 Free Software Foundation, Inc.
3
4 This file is part of GNU CC.
5
6 GNU CC is free software; you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation; either version 2, or (at your option)
9 any later version.
10
11 GNU CC is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
15
16 You should have received a copy of the GNU General Public License
17 along with GNU CC; see the file COPYING. If not, write to
18 the Free Software Foundation, 59 Temple Place - Suite 330,
19 Boston, MA 02111-1307, USA. */
20
21
22 /* This is the loop optimization pass of the compiler.
23 It finds invariant computations within loops and moves them
24 to the beginning of the loop. Then it identifies basic and
25 general induction variables. Strength reduction is applied to the general
26 induction variables, and induction variable elimination is applied to
27 the basic induction variables.
28
29 It also finds cases where
30 a register is set within the loop by zero-extending a narrower value
31 and changes these to zero the entire register once before the loop
32 and merely copy the low part within the loop.
33
34 Most of the complexity is in heuristics to decide when it is worth
35 while to do these things. */
36
37 #include <stdio.h>
38 #include "config.h"
39 #include "rtl.h"
40 #include "obstack.h"
41 #include "expr.h"
42 #include "insn-config.h"
43 #include "insn-flags.h"
44 #include "regs.h"
45 #include "hard-reg-set.h"
46 #include "recog.h"
47 #include "flags.h"
48 #include "real.h"
49 #include "loop.h"
50 #include "except.h"
51
52 /* Vector mapping INSN_UIDs to luids.
53 The luids are like uids but increase monotonically always.
54 We use them to see whether a jump comes from outside a given loop. */
55
56 int *uid_luid;
57
58 /* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop
59 number the insn is contained in. */
60
61 int *uid_loop_num;
62
63 /* 1 + largest uid of any insn. */
64
65 int max_uid_for_loop;
66
67 /* 1 + luid of last insn. */
68
69 static int max_luid;
70
71 /* Number of loops detected in current function. Used as index to the
72 next few tables. */
73
74 static int max_loop_num;
75
76 /* Indexed by loop number, contains the first and last insn of each loop. */
77
78 static rtx *loop_number_loop_starts, *loop_number_loop_ends;
79
80 /* For each loop, gives the containing loop number, -1 if none. */
81
82 int *loop_outer_loop;
83
84 /* Indexed by loop number, contains a nonzero value if the "loop" isn't
85 really a loop (an insn outside the loop branches into it). */
86
87 static char *loop_invalid;
88
89 /* Indexed by loop number, links together all LABEL_REFs which refer to
90 code labels outside the loop. Used by routines that need to know all
91 loop exits, such as final_biv_value and final_giv_value.
92
93 This does not include loop exits due to return instructions. This is
94 because all bivs and givs are pseudos, and hence must be dead after a
95 return, so the presense of a return does not affect any of the
96 optimizations that use this info. It is simpler to just not include return
97 instructions on this list. */
98
99 rtx *loop_number_exit_labels;
100
101 /* Indexed by loop number, counts the number of LABEL_REFs on
102 loop_number_exit_labels for this loop and all loops nested inside it. */
103
104 int *loop_number_exit_count;
105
106 /* Holds the number of loop iterations. It is zero if the number could not be
107 calculated. Must be unsigned since the number of iterations can
108 be as high as 2^wordsize-1. For loops with a wider iterator, this number
109 will will be zero if the number of loop iterations is too large for an
110 unsigned integer to hold. */
111
112 unsigned HOST_WIDE_INT loop_n_iterations;
113
114 /* Nonzero if there is a subroutine call in the current loop.
115 (unknown_address_altered is also nonzero in this case.) */
116
117 static int loop_has_call;
118
119 /* Nonzero if there is a volatile memory reference in the current
120 loop. */
121
122 static int loop_has_volatile;
123
124 /* Added loop_continue which is the NOTE_INSN_LOOP_CONT of the
125 current loop. A continue statement will generate a branch to
126 NEXT_INSN (loop_continue). */
127
128 static rtx loop_continue;
129
130 /* Indexed by register number, contains the number of times the reg
131 is set during the loop being scanned.
132 During code motion, a negative value indicates a reg that has been
133 made a candidate; in particular -2 means that it is an candidate that
134 we know is equal to a constant and -1 means that it is an candidate
135 not known equal to a constant.
136 After code motion, regs moved have 0 (which is accurate now)
137 while the failed candidates have the original number of times set.
138
139 Therefore, at all times, == 0 indicates an invariant register;
140 < 0 a conditionally invariant one. */
141
142 static short *n_times_set;
143
144 /* Original value of n_times_set; same except that this value
145 is not set negative for a reg whose sets have been made candidates
146 and not set to 0 for a reg that is moved. */
147
148 static short *n_times_used;
149
150 /* Index by register number, 1 indicates that the register
151 cannot be moved or strength reduced. */
152
153 static char *may_not_optimize;
154
155 /* Nonzero means reg N has already been moved out of one loop.
156 This reduces the desire to move it out of another. */
157
158 static char *moved_once;
159
160 /* Array of MEMs that are stored in this loop. If there are too many to fit
161 here, we just turn on unknown_address_altered. */
162
163 #define NUM_STORES 20
164 static rtx loop_store_mems[NUM_STORES];
165
166 /* Index of first available slot in above array. */
167 static int loop_store_mems_idx;
168
169 /* Nonzero if we don't know what MEMs were changed in the current loop.
170 This happens if the loop contains a call (in which case `loop_has_call'
171 will also be set) or if we store into more than NUM_STORES MEMs. */
172
173 static int unknown_address_altered;
174
175 /* Count of movable (i.e. invariant) instructions discovered in the loop. */
176 static int num_movables;
177
178 /* Count of memory write instructions discovered in the loop. */
179 static int num_mem_sets;
180
181 /* Number of loops contained within the current one, including itself. */
182 static int loops_enclosed;
183
184 /* Bound on pseudo register number before loop optimization.
185 A pseudo has valid regscan info if its number is < max_reg_before_loop. */
186 int max_reg_before_loop;
187
188 /* This obstack is used in product_cheap_p to allocate its rtl. It
189 may call gen_reg_rtx which, in turn, may reallocate regno_reg_rtx.
190 If we used the same obstack that it did, we would be deallocating
191 that array. */
192
193 static struct obstack temp_obstack;
194
195 /* This is where the pointer to the obstack being used for RTL is stored. */
196
197 extern struct obstack *rtl_obstack;
198
199 #define obstack_chunk_alloc xmalloc
200 #define obstack_chunk_free free
201
202 extern char *oballoc ();
203 \f
204 /* During the analysis of a loop, a chain of `struct movable's
205 is made to record all the movable insns found.
206 Then the entire chain can be scanned to decide which to move. */
207
208 struct movable
209 {
210 rtx insn; /* A movable insn */
211 rtx set_src; /* The expression this reg is set from. */
212 rtx set_dest; /* The destination of this SET. */
213 rtx dependencies; /* When INSN is libcall, this is an EXPR_LIST
214 of any registers used within the LIBCALL. */
215 int consec; /* Number of consecutive following insns
216 that must be moved with this one. */
217 int regno; /* The register it sets */
218 short lifetime; /* lifetime of that register;
219 may be adjusted when matching movables
220 that load the same value are found. */
221 short savings; /* Number of insns we can move for this reg,
222 including other movables that force this
223 or match this one. */
224 unsigned int cond : 1; /* 1 if only conditionally movable */
225 unsigned int force : 1; /* 1 means MUST move this insn */
226 unsigned int global : 1; /* 1 means reg is live outside this loop */
227 /* If PARTIAL is 1, GLOBAL means something different:
228 that the reg is live outside the range from where it is set
229 to the following label. */
230 unsigned int done : 1; /* 1 inhibits further processing of this */
231
232 unsigned int partial : 1; /* 1 means this reg is used for zero-extending.
233 In particular, moving it does not make it
234 invariant. */
235 unsigned int move_insn : 1; /* 1 means that we call emit_move_insn to
236 load SRC, rather than copying INSN. */
237 unsigned int is_equiv : 1; /* 1 means a REG_EQUIV is present on INSN. */
238 enum machine_mode savemode; /* Nonzero means it is a mode for a low part
239 that we should avoid changing when clearing
240 the rest of the reg. */
241 struct movable *match; /* First entry for same value */
242 struct movable *forces; /* An insn that must be moved if this is */
243 struct movable *next;
244 };
245
246 FILE *loop_dump_stream;
247
248 /* Forward declarations. */
249
250 static void find_and_verify_loops ();
251 static void mark_loop_jump ();
252 static void prescan_loop ();
253 static int reg_in_basic_block_p ();
254 static int consec_sets_invariant_p ();
255 static rtx libcall_other_reg ();
256 static int labels_in_range_p ();
257 static void count_loop_regs_set ();
258 static void note_addr_stored ();
259 static int loop_reg_used_before_p ();
260 static void scan_loop ();
261 static void replace_call_address ();
262 static rtx skip_consec_insns ();
263 static int libcall_benefit ();
264 static void ignore_some_movables ();
265 static void force_movables ();
266 static void combine_movables ();
267 static int rtx_equal_for_loop_p ();
268 static void move_movables ();
269 static void strength_reduce ();
270 static int valid_initial_value_p ();
271 static void find_mem_givs ();
272 static void record_biv ();
273 static void check_final_value ();
274 static void record_giv ();
275 static void update_giv_derive ();
276 static int basic_induction_var ();
277 static rtx simplify_giv_expr ();
278 static int general_induction_var ();
279 static int consec_sets_giv ();
280 static int check_dbra_loop ();
281 static rtx express_from ();
282 static int combine_givs_p ();
283 static void combine_givs ();
284 static int product_cheap_p ();
285 static int maybe_eliminate_biv ();
286 static int maybe_eliminate_biv_1 ();
287 static int last_use_this_basic_block ();
288 static void record_initial ();
289 static void update_reg_last_use ();
290 \f
291 /* Relative gain of eliminating various kinds of operations. */
292 int add_cost;
293 #if 0
294 int shift_cost;
295 int mult_cost;
296 #endif
297
298 /* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to
299 copy the value of the strength reduced giv to its original register. */
300 int copy_cost;
301
302 void
303 init_loop ()
304 {
305 char *free_point = (char *) oballoc (1);
306 rtx reg = gen_rtx (REG, word_mode, LAST_VIRTUAL_REGISTER + 1);
307
308 add_cost = rtx_cost (gen_rtx (PLUS, word_mode, reg, reg), SET);
309
310 /* We multiply by 2 to reconcile the difference in scale between
311 these two ways of computing costs. Otherwise the cost of a copy
312 will be far less than the cost of an add. */
313
314 copy_cost = 2 * 2;
315
316 /* Free the objects we just allocated. */
317 obfree (free_point);
318
319 /* Initialize the obstack used for rtl in product_cheap_p. */
320 gcc_obstack_init (&temp_obstack);
321 }
322 \f
323 /* Entry point of this file. Perform loop optimization
324 on the current function. F is the first insn of the function
325 and DUMPFILE is a stream for output of a trace of actions taken
326 (or 0 if none should be output). */
327
328 void
329 loop_optimize (f, dumpfile)
330 /* f is the first instruction of a chain of insns for one function */
331 rtx f;
332 FILE *dumpfile;
333 {
334 register rtx insn;
335 register int i;
336 rtx last_insn;
337
338 loop_dump_stream = dumpfile;
339
340 init_recog_no_volatile ();
341 init_alias_analysis ();
342
343 max_reg_before_loop = max_reg_num ();
344
345 moved_once = (char *) alloca (max_reg_before_loop);
346 bzero (moved_once, max_reg_before_loop);
347
348 regs_may_share = 0;
349
350 /* Count the number of loops. */
351
352 max_loop_num = 0;
353 for (insn = f; insn; insn = NEXT_INSN (insn))
354 {
355 if (GET_CODE (insn) == NOTE
356 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
357 max_loop_num++;
358 }
359
360 /* Don't waste time if no loops. */
361 if (max_loop_num == 0)
362 return;
363
364 /* Get size to use for tables indexed by uids.
365 Leave some space for labels allocated by find_and_verify_loops. */
366 max_uid_for_loop = get_max_uid () + 1 + max_loop_num * 32;
367
368 uid_luid = (int *) alloca (max_uid_for_loop * sizeof (int));
369 uid_loop_num = (int *) alloca (max_uid_for_loop * sizeof (int));
370
371 bzero ((char *) uid_luid, max_uid_for_loop * sizeof (int));
372 bzero ((char *) uid_loop_num, max_uid_for_loop * sizeof (int));
373
374 /* Allocate tables for recording each loop. We set each entry, so they need
375 not be zeroed. */
376 loop_number_loop_starts = (rtx *) alloca (max_loop_num * sizeof (rtx));
377 loop_number_loop_ends = (rtx *) alloca (max_loop_num * sizeof (rtx));
378 loop_outer_loop = (int *) alloca (max_loop_num * sizeof (int));
379 loop_invalid = (char *) alloca (max_loop_num * sizeof (char));
380 loop_number_exit_labels = (rtx *) alloca (max_loop_num * sizeof (rtx));
381 loop_number_exit_count = (int *) alloca (max_loop_num * sizeof (int));
382
383 /* Find and process each loop.
384 First, find them, and record them in order of their beginnings. */
385 find_and_verify_loops (f);
386
387 /* Now find all register lifetimes. This must be done after
388 find_and_verify_loops, because it might reorder the insns in the
389 function. */
390 reg_scan (f, max_reg_num (), 1);
391
392 /* See if we went too far. */
393 if (get_max_uid () > max_uid_for_loop)
394 abort ();
395
396 /* Compute the mapping from uids to luids.
397 LUIDs are numbers assigned to insns, like uids,
398 except that luids increase monotonically through the code.
399 Don't assign luids to line-number NOTEs, so that the distance in luids
400 between two insns is not affected by -g. */
401
402 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
403 {
404 last_insn = insn;
405 if (GET_CODE (insn) != NOTE
406 || NOTE_LINE_NUMBER (insn) <= 0)
407 uid_luid[INSN_UID (insn)] = ++i;
408 else
409 /* Give a line number note the same luid as preceding insn. */
410 uid_luid[INSN_UID (insn)] = i;
411 }
412
413 max_luid = i + 1;
414
415 /* Don't leave gaps in uid_luid for insns that have been
416 deleted. It is possible that the first or last insn
417 using some register has been deleted by cross-jumping.
418 Make sure that uid_luid for that former insn's uid
419 points to the general area where that insn used to be. */
420 for (i = 0; i < max_uid_for_loop; i++)
421 {
422 uid_luid[0] = uid_luid[i];
423 if (uid_luid[0] != 0)
424 break;
425 }
426 for (i = 0; i < max_uid_for_loop; i++)
427 if (uid_luid[i] == 0)
428 uid_luid[i] = uid_luid[i - 1];
429
430 /* Create a mapping from loops to BLOCK tree nodes. */
431 if (flag_unroll_loops && write_symbols != NO_DEBUG)
432 find_loop_tree_blocks ();
433
434 /* Now scan the loops, last ones first, since this means inner ones are done
435 before outer ones. */
436 for (i = max_loop_num-1; i >= 0; i--)
437 if (! loop_invalid[i] && loop_number_loop_ends[i])
438 scan_loop (loop_number_loop_starts[i], loop_number_loop_ends[i],
439 max_reg_num ());
440
441 /* If debugging and unrolling loops, we must replicate the tree nodes
442 corresponding to the blocks inside the loop, so that the original one
443 to one mapping will remain. */
444 if (flag_unroll_loops && write_symbols != NO_DEBUG)
445 unroll_block_trees ();
446 }
447 \f
448 /* Optimize one loop whose start is LOOP_START and end is END.
449 LOOP_START is the NOTE_INSN_LOOP_BEG and END is the matching
450 NOTE_INSN_LOOP_END. */
451
452 /* ??? Could also move memory writes out of loops if the destination address
453 is invariant, the source is invariant, the memory write is not volatile,
454 and if we can prove that no read inside the loop can read this address
455 before the write occurs. If there is a read of this address after the
456 write, then we can also mark the memory read as invariant. */
457
458 static void
459 scan_loop (loop_start, end, nregs)
460 rtx loop_start, end;
461 int nregs;
462 {
463 register int i;
464 register rtx p;
465 /* 1 if we are scanning insns that could be executed zero times. */
466 int maybe_never = 0;
467 /* 1 if we are scanning insns that might never be executed
468 due to a subroutine call which might exit before they are reached. */
469 int call_passed = 0;
470 /* For a rotated loop that is entered near the bottom,
471 this is the label at the top. Otherwise it is zero. */
472 rtx loop_top = 0;
473 /* Jump insn that enters the loop, or 0 if control drops in. */
474 rtx loop_entry_jump = 0;
475 /* Place in the loop where control enters. */
476 rtx scan_start;
477 /* Number of insns in the loop. */
478 int insn_count;
479 int in_libcall = 0;
480 int tem;
481 rtx temp;
482 /* The SET from an insn, if it is the only SET in the insn. */
483 rtx set, set1;
484 /* Chain describing insns movable in current loop. */
485 struct movable *movables = 0;
486 /* Last element in `movables' -- so we can add elements at the end. */
487 struct movable *last_movable = 0;
488 /* Ratio of extra register life span we can justify
489 for saving an instruction. More if loop doesn't call subroutines
490 since in that case saving an insn makes more difference
491 and more registers are available. */
492 int threshold;
493 /* If we have calls, contains the insn in which a register was used
494 if it was used exactly once; contains const0_rtx if it was used more
495 than once. */
496 rtx *reg_single_usage = 0;
497 /* Nonzero if we are scanning instructions in a sub-loop. */
498 int loop_depth = 0;
499
500 n_times_set = (short *) alloca (nregs * sizeof (short));
501 n_times_used = (short *) alloca (nregs * sizeof (short));
502 may_not_optimize = (char *) alloca (nregs);
503
504 /* Determine whether this loop starts with a jump down to a test at
505 the end. This will occur for a small number of loops with a test
506 that is too complex to duplicate in front of the loop.
507
508 We search for the first insn or label in the loop, skipping NOTEs.
509 However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG
510 (because we might have a loop executed only once that contains a
511 loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END
512 (in case we have a degenerate loop).
513
514 Note that if we mistakenly think that a loop is entered at the top
515 when, in fact, it is entered at the exit test, the only effect will be
516 slightly poorer optimization. Making the opposite error can generate
517 incorrect code. Since very few loops now start with a jump to the
518 exit test, the code here to detect that case is very conservative. */
519
520 for (p = NEXT_INSN (loop_start);
521 p != end
522 && GET_CODE (p) != CODE_LABEL && GET_RTX_CLASS (GET_CODE (p)) != 'i'
523 && (GET_CODE (p) != NOTE
524 || (NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_BEG
525 && NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_END));
526 p = NEXT_INSN (p))
527 ;
528
529 scan_start = p;
530
531 /* Set up variables describing this loop. */
532 prescan_loop (loop_start, end);
533 threshold = (loop_has_call ? 1 : 2) * (1 + n_non_fixed_regs);
534
535 /* If loop has a jump before the first label,
536 the true entry is the target of that jump.
537 Start scan from there.
538 But record in LOOP_TOP the place where the end-test jumps
539 back to so we can scan that after the end of the loop. */
540 if (GET_CODE (p) == JUMP_INSN)
541 {
542 loop_entry_jump = p;
543
544 /* Loop entry must be unconditional jump (and not a RETURN) */
545 if (simplejump_p (p)
546 && JUMP_LABEL (p) != 0
547 /* Check to see whether the jump actually
548 jumps out of the loop (meaning it's no loop).
549 This case can happen for things like
550 do {..} while (0). If this label was generated previously
551 by loop, we can't tell anything about it and have to reject
552 the loop. */
553 && INSN_UID (JUMP_LABEL (p)) < max_uid_for_loop
554 && INSN_LUID (JUMP_LABEL (p)) >= INSN_LUID (loop_start)
555 && INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (end))
556 {
557 loop_top = next_label (scan_start);
558 scan_start = JUMP_LABEL (p);
559 }
560 }
561
562 /* If SCAN_START was an insn created by loop, we don't know its luid
563 as required by loop_reg_used_before_p. So skip such loops. (This
564 test may never be true, but it's best to play it safe.)
565
566 Also, skip loops where we do not start scanning at a label. This
567 test also rejects loops starting with a JUMP_INSN that failed the
568 test above. */
569
570 if (INSN_UID (scan_start) >= max_uid_for_loop
571 || GET_CODE (scan_start) != CODE_LABEL)
572 {
573 if (loop_dump_stream)
574 fprintf (loop_dump_stream, "\nLoop from %d to %d is phony.\n\n",
575 INSN_UID (loop_start), INSN_UID (end));
576 return;
577 }
578
579 /* Count number of times each reg is set during this loop.
580 Set may_not_optimize[I] if it is not safe to move out
581 the setting of register I. If this loop has calls, set
582 reg_single_usage[I]. */
583
584 bzero ((char *) n_times_set, nregs * sizeof (short));
585 bzero (may_not_optimize, nregs);
586
587 if (loop_has_call)
588 {
589 reg_single_usage = (rtx *) alloca (nregs * sizeof (rtx));
590 bzero ((char *) reg_single_usage, nregs * sizeof (rtx));
591 }
592
593 count_loop_regs_set (loop_top ? loop_top : loop_start, end,
594 may_not_optimize, reg_single_usage, &insn_count, nregs);
595
596 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
597 may_not_optimize[i] = 1, n_times_set[i] = 1;
598 bcopy ((char *) n_times_set, (char *) n_times_used, nregs * sizeof (short));
599
600 if (loop_dump_stream)
601 {
602 fprintf (loop_dump_stream, "\nLoop from %d to %d: %d real insns.\n",
603 INSN_UID (loop_start), INSN_UID (end), insn_count);
604 if (loop_continue)
605 fprintf (loop_dump_stream, "Continue at insn %d.\n",
606 INSN_UID (loop_continue));
607 }
608
609 /* Scan through the loop finding insns that are safe to move.
610 Set n_times_set negative for the reg being set, so that
611 this reg will be considered invariant for subsequent insns.
612 We consider whether subsequent insns use the reg
613 in deciding whether it is worth actually moving.
614
615 MAYBE_NEVER is nonzero if we have passed a conditional jump insn
616 and therefore it is possible that the insns we are scanning
617 would never be executed. At such times, we must make sure
618 that it is safe to execute the insn once instead of zero times.
619 When MAYBE_NEVER is 0, all insns will be executed at least once
620 so that is not a problem. */
621
622 p = scan_start;
623 while (1)
624 {
625 p = NEXT_INSN (p);
626 /* At end of a straight-in loop, we are done.
627 At end of a loop entered at the bottom, scan the top. */
628 if (p == scan_start)
629 break;
630 if (p == end)
631 {
632 if (loop_top != 0)
633 p = loop_top;
634 else
635 break;
636 if (p == scan_start)
637 break;
638 }
639
640 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
641 && find_reg_note (p, REG_LIBCALL, NULL_RTX))
642 in_libcall = 1;
643 else if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
644 && find_reg_note (p, REG_RETVAL, NULL_RTX))
645 in_libcall = 0;
646
647 if (GET_CODE (p) == INSN
648 && (set = single_set (p))
649 && GET_CODE (SET_DEST (set)) == REG
650 && ! may_not_optimize[REGNO (SET_DEST (set))])
651 {
652 int tem1 = 0;
653 int tem2 = 0;
654 int move_insn = 0;
655 rtx src = SET_SRC (set);
656 rtx dependencies = 0;
657
658 /* Figure out what to use as a source of this insn. If a REG_EQUIV
659 note is given or if a REG_EQUAL note with a constant operand is
660 specified, use it as the source and mark that we should move
661 this insn by calling emit_move_insn rather that duplicating the
662 insn.
663
664 Otherwise, only use the REG_EQUAL contents if a REG_RETVAL note
665 is present. */
666 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
667 if (temp)
668 src = XEXP (temp, 0), move_insn = 1;
669 else
670 {
671 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
672 if (temp && CONSTANT_P (XEXP (temp, 0)))
673 src = XEXP (temp, 0), move_insn = 1;
674 if (temp && find_reg_note (p, REG_RETVAL, NULL_RTX))
675 {
676 src = XEXP (temp, 0);
677 /* A libcall block can use regs that don't appear in
678 the equivalent expression. To move the libcall,
679 we must move those regs too. */
680 dependencies = libcall_other_reg (p, src);
681 }
682 }
683
684 /* Don't try to optimize a register that was made
685 by loop-optimization for an inner loop.
686 We don't know its life-span, so we can't compute the benefit. */
687 if (REGNO (SET_DEST (set)) >= max_reg_before_loop)
688 ;
689 /* In order to move a register, we need to have one of three cases:
690 (1) it is used only in the same basic block as the set
691 (2) it is not a user variable and it is not used in the
692 exit test (this can cause the variable to be used
693 before it is set just like a user-variable).
694 (3) the set is guaranteed to be executed once the loop starts,
695 and the reg is not used until after that. */
696 else if (! ((! maybe_never
697 && ! loop_reg_used_before_p (set, p, loop_start,
698 scan_start, end))
699 || (! REG_USERVAR_P (SET_DEST (set))
700 && ! REG_LOOP_TEST_P (SET_DEST (set)))
701 || reg_in_basic_block_p (p, SET_DEST (set))))
702 ;
703 else if ((tem = invariant_p (src))
704 && (dependencies == 0
705 || (tem2 = invariant_p (dependencies)) != 0)
706 && (n_times_set[REGNO (SET_DEST (set))] == 1
707 || (tem1
708 = consec_sets_invariant_p (SET_DEST (set),
709 n_times_set[REGNO (SET_DEST (set))],
710 p)))
711 /* If the insn can cause a trap (such as divide by zero),
712 can't move it unless it's guaranteed to be executed
713 once loop is entered. Even a function call might
714 prevent the trap insn from being reached
715 (since it might exit!) */
716 && ! ((maybe_never || call_passed)
717 && may_trap_p (src)))
718 {
719 register struct movable *m;
720 register int regno = REGNO (SET_DEST (set));
721
722 /* A potential lossage is where we have a case where two insns
723 can be combined as long as they are both in the loop, but
724 we move one of them outside the loop. For large loops,
725 this can lose. The most common case of this is the address
726 of a function being called.
727
728 Therefore, if this register is marked as being used exactly
729 once if we are in a loop with calls (a "large loop"), see if
730 we can replace the usage of this register with the source
731 of this SET. If we can, delete this insn.
732
733 Don't do this if P has a REG_RETVAL note or if we have
734 SMALL_REGISTER_CLASSES and SET_SRC is a hard register. */
735
736 if (reg_single_usage && reg_single_usage[regno] != 0
737 && reg_single_usage[regno] != const0_rtx
738 && regno_first_uid[regno] == INSN_UID (p)
739 && (regno_last_uid[regno]
740 == INSN_UID (reg_single_usage[regno]))
741 && n_times_set[REGNO (SET_DEST (set))] == 1
742 && ! side_effects_p (SET_SRC (set))
743 && ! find_reg_note (p, REG_RETVAL, NULL_RTX)
744 #ifdef SMALL_REGISTER_CLASSES
745 && ! (GET_CODE (SET_SRC (set)) == REG
746 && REGNO (SET_SRC (set)) < FIRST_PSEUDO_REGISTER)
747 #endif
748 /* This test is not redundant; SET_SRC (set) might be
749 a call-clobbered register and the life of REGNO
750 might span a call. */
751 && ! modified_between_p (SET_SRC (set), p,
752 reg_single_usage[regno])
753 && no_labels_between_p (p, reg_single_usage[regno])
754 && validate_replace_rtx (SET_DEST (set), SET_SRC (set),
755 reg_single_usage[regno]))
756 {
757 /* Replace any usage in a REG_EQUAL note. Must copy the
758 new source, so that we don't get rtx sharing between the
759 SET_SOURCE and REG_NOTES of insn p. */
760 REG_NOTES (reg_single_usage[regno])
761 = replace_rtx (REG_NOTES (reg_single_usage[regno]),
762 SET_DEST (set), copy_rtx (SET_SRC (set)));
763
764 PUT_CODE (p, NOTE);
765 NOTE_LINE_NUMBER (p) = NOTE_INSN_DELETED;
766 NOTE_SOURCE_FILE (p) = 0;
767 n_times_set[regno] = 0;
768 continue;
769 }
770
771 m = (struct movable *) alloca (sizeof (struct movable));
772 m->next = 0;
773 m->insn = p;
774 m->set_src = src;
775 m->dependencies = dependencies;
776 m->set_dest = SET_DEST (set);
777 m->force = 0;
778 m->consec = n_times_set[REGNO (SET_DEST (set))] - 1;
779 m->done = 0;
780 m->forces = 0;
781 m->partial = 0;
782 m->move_insn = move_insn;
783 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
784 m->savemode = VOIDmode;
785 m->regno = regno;
786 /* Set M->cond if either invariant_p or consec_sets_invariant_p
787 returned 2 (only conditionally invariant). */
788 m->cond = ((tem | tem1 | tem2) > 1);
789 m->global = (uid_luid[regno_last_uid[regno]] > INSN_LUID (end)
790 || uid_luid[regno_first_uid[regno]] < INSN_LUID (loop_start));
791 m->match = 0;
792 m->lifetime = (uid_luid[regno_last_uid[regno]]
793 - uid_luid[regno_first_uid[regno]]);
794 m->savings = n_times_used[regno];
795 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
796 m->savings += libcall_benefit (p);
797 n_times_set[regno] = move_insn ? -2 : -1;
798 /* Add M to the end of the chain MOVABLES. */
799 if (movables == 0)
800 movables = m;
801 else
802 last_movable->next = m;
803 last_movable = m;
804
805 if (m->consec > 0)
806 {
807 /* Skip this insn, not checking REG_LIBCALL notes. */
808 p = next_nonnote_insn (p);
809 /* Skip the consecutive insns, if there are any. */
810 p = skip_consec_insns (p, m->consec);
811 /* Back up to the last insn of the consecutive group. */
812 p = prev_nonnote_insn (p);
813
814 /* We must now reset m->move_insn, m->is_equiv, and possibly
815 m->set_src to correspond to the effects of all the
816 insns. */
817 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
818 if (temp)
819 m->set_src = XEXP (temp, 0), m->move_insn = 1;
820 else
821 {
822 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
823 if (temp && CONSTANT_P (XEXP (temp, 0)))
824 m->set_src = XEXP (temp, 0), m->move_insn = 1;
825 else
826 m->move_insn = 0;
827
828 }
829 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
830 }
831 }
832 /* If this register is always set within a STRICT_LOW_PART
833 or set to zero, then its high bytes are constant.
834 So clear them outside the loop and within the loop
835 just load the low bytes.
836 We must check that the machine has an instruction to do so.
837 Also, if the value loaded into the register
838 depends on the same register, this cannot be done. */
839 else if (SET_SRC (set) == const0_rtx
840 && GET_CODE (NEXT_INSN (p)) == INSN
841 && (set1 = single_set (NEXT_INSN (p)))
842 && GET_CODE (set1) == SET
843 && (GET_CODE (SET_DEST (set1)) == STRICT_LOW_PART)
844 && (GET_CODE (XEXP (SET_DEST (set1), 0)) == SUBREG)
845 && (SUBREG_REG (XEXP (SET_DEST (set1), 0))
846 == SET_DEST (set))
847 && !reg_mentioned_p (SET_DEST (set), SET_SRC (set1)))
848 {
849 register int regno = REGNO (SET_DEST (set));
850 if (n_times_set[regno] == 2)
851 {
852 register struct movable *m;
853 m = (struct movable *) alloca (sizeof (struct movable));
854 m->next = 0;
855 m->insn = p;
856 m->set_dest = SET_DEST (set);
857 m->dependencies = 0;
858 m->force = 0;
859 m->consec = 0;
860 m->done = 0;
861 m->forces = 0;
862 m->move_insn = 0;
863 m->partial = 1;
864 /* If the insn may not be executed on some cycles,
865 we can't clear the whole reg; clear just high part.
866 Not even if the reg is used only within this loop.
867 Consider this:
868 while (1)
869 while (s != t) {
870 if (foo ()) x = *s;
871 use (x);
872 }
873 Clearing x before the inner loop could clobber a value
874 being saved from the last time around the outer loop.
875 However, if the reg is not used outside this loop
876 and all uses of the register are in the same
877 basic block as the store, there is no problem.
878
879 If this insn was made by loop, we don't know its
880 INSN_LUID and hence must make a conservative
881 assumption. */
882 m->global = (INSN_UID (p) >= max_uid_for_loop
883 || (uid_luid[regno_last_uid[regno]]
884 > INSN_LUID (end))
885 || (uid_luid[regno_first_uid[regno]]
886 < INSN_LUID (p))
887 || (labels_in_range_p
888 (p, uid_luid[regno_first_uid[regno]])));
889 if (maybe_never && m->global)
890 m->savemode = GET_MODE (SET_SRC (set1));
891 else
892 m->savemode = VOIDmode;
893 m->regno = regno;
894 m->cond = 0;
895 m->match = 0;
896 m->lifetime = (uid_luid[regno_last_uid[regno]]
897 - uid_luid[regno_first_uid[regno]]);
898 m->savings = 1;
899 n_times_set[regno] = -1;
900 /* Add M to the end of the chain MOVABLES. */
901 if (movables == 0)
902 movables = m;
903 else
904 last_movable->next = m;
905 last_movable = m;
906 }
907 }
908 }
909 /* Past a call insn, we get to insns which might not be executed
910 because the call might exit. This matters for insns that trap.
911 Call insns inside a REG_LIBCALL/REG_RETVAL block always return,
912 so they don't count. */
913 else if (GET_CODE (p) == CALL_INSN && ! in_libcall)
914 call_passed = 1;
915 /* Past a label or a jump, we get to insns for which we
916 can't count on whether or how many times they will be
917 executed during each iteration. Therefore, we can
918 only move out sets of trivial variables
919 (those not used after the loop). */
920 /* Similar code appears twice in strength_reduce. */
921 else if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
922 /* If we enter the loop in the middle, and scan around to the
923 beginning, don't set maybe_never for that. This must be an
924 unconditional jump, otherwise the code at the top of the
925 loop might never be executed. Unconditional jumps are
926 followed a by barrier then loop end. */
927 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop_top
928 && NEXT_INSN (NEXT_INSN (p)) == end
929 && simplejump_p (p)))
930 maybe_never = 1;
931 else if (GET_CODE (p) == NOTE)
932 {
933 /* At the virtual top of a converted loop, insns are again known to
934 be executed: logically, the loop begins here even though the exit
935 code has been duplicated. */
936 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
937 maybe_never = call_passed = 0;
938 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
939 loop_depth++;
940 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
941 loop_depth--;
942 }
943 }
944
945 /* If one movable subsumes another, ignore that other. */
946
947 ignore_some_movables (movables);
948
949 /* For each movable insn, see if the reg that it loads
950 leads when it dies right into another conditionally movable insn.
951 If so, record that the second insn "forces" the first one,
952 since the second can be moved only if the first is. */
953
954 force_movables (movables);
955
956 /* See if there are multiple movable insns that load the same value.
957 If there are, make all but the first point at the first one
958 through the `match' field, and add the priorities of them
959 all together as the priority of the first. */
960
961 combine_movables (movables, nregs);
962
963 /* Now consider each movable insn to decide whether it is worth moving.
964 Store 0 in n_times_set for each reg that is moved. */
965
966 move_movables (movables, threshold,
967 insn_count, loop_start, end, nregs);
968
969 /* Now candidates that still are negative are those not moved.
970 Change n_times_set to indicate that those are not actually invariant. */
971 for (i = 0; i < nregs; i++)
972 if (n_times_set[i] < 0)
973 n_times_set[i] = n_times_used[i];
974
975 if (flag_strength_reduce)
976 strength_reduce (scan_start, end, loop_top,
977 insn_count, loop_start, end);
978 }
979 \f
980 /* Add elements to *OUTPUT to record all the pseudo-regs
981 mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */
982
983 void
984 record_excess_regs (in_this, not_in_this, output)
985 rtx in_this, not_in_this;
986 rtx *output;
987 {
988 enum rtx_code code;
989 char *fmt;
990 int i;
991
992 code = GET_CODE (in_this);
993
994 switch (code)
995 {
996 case PC:
997 case CC0:
998 case CONST_INT:
999 case CONST_DOUBLE:
1000 case CONST:
1001 case SYMBOL_REF:
1002 case LABEL_REF:
1003 return;
1004
1005 case REG:
1006 if (REGNO (in_this) >= FIRST_PSEUDO_REGISTER
1007 && ! reg_mentioned_p (in_this, not_in_this))
1008 *output = gen_rtx (EXPR_LIST, VOIDmode, in_this, *output);
1009 return;
1010 }
1011
1012 fmt = GET_RTX_FORMAT (code);
1013 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1014 {
1015 int j;
1016
1017 switch (fmt[i])
1018 {
1019 case 'E':
1020 for (j = 0; j < XVECLEN (in_this, i); j++)
1021 record_excess_regs (XVECEXP (in_this, i, j), not_in_this, output);
1022 break;
1023
1024 case 'e':
1025 record_excess_regs (XEXP (in_this, i), not_in_this, output);
1026 break;
1027 }
1028 }
1029 }
1030 \f
1031 /* Check what regs are referred to in the libcall block ending with INSN,
1032 aside from those mentioned in the equivalent value.
1033 If there are none, return 0.
1034 If there are one or more, return an EXPR_LIST containing all of them. */
1035
1036 static rtx
1037 libcall_other_reg (insn, equiv)
1038 rtx insn, equiv;
1039 {
1040 rtx note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
1041 rtx p = XEXP (note, 0);
1042 rtx output = 0;
1043
1044 /* First, find all the regs used in the libcall block
1045 that are not mentioned as inputs to the result. */
1046
1047 while (p != insn)
1048 {
1049 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1050 || GET_CODE (p) == CALL_INSN)
1051 record_excess_regs (PATTERN (p), equiv, &output);
1052 p = NEXT_INSN (p);
1053 }
1054
1055 return output;
1056 }
1057 \f
1058 /* Return 1 if all uses of REG
1059 are between INSN and the end of the basic block. */
1060
1061 static int
1062 reg_in_basic_block_p (insn, reg)
1063 rtx insn, reg;
1064 {
1065 int regno = REGNO (reg);
1066 rtx p;
1067
1068 if (regno_first_uid[regno] != INSN_UID (insn))
1069 return 0;
1070
1071 /* Search this basic block for the already recorded last use of the reg. */
1072 for (p = insn; p; p = NEXT_INSN (p))
1073 {
1074 switch (GET_CODE (p))
1075 {
1076 case NOTE:
1077 break;
1078
1079 case INSN:
1080 case CALL_INSN:
1081 /* Ordinary insn: if this is the last use, we win. */
1082 if (regno_last_uid[regno] == INSN_UID (p))
1083 return 1;
1084 break;
1085
1086 case JUMP_INSN:
1087 /* Jump insn: if this is the last use, we win. */
1088 if (regno_last_uid[regno] == INSN_UID (p))
1089 return 1;
1090 /* Otherwise, it's the end of the basic block, so we lose. */
1091 return 0;
1092
1093 case CODE_LABEL:
1094 case BARRIER:
1095 /* It's the end of the basic block, so we lose. */
1096 return 0;
1097 }
1098 }
1099
1100 /* The "last use" doesn't follow the "first use"?? */
1101 abort ();
1102 }
1103 \f
1104 /* Compute the benefit of eliminating the insns in the block whose
1105 last insn is LAST. This may be a group of insns used to compute a
1106 value directly or can contain a library call. */
1107
1108 static int
1109 libcall_benefit (last)
1110 rtx last;
1111 {
1112 rtx insn;
1113 int benefit = 0;
1114
1115 for (insn = XEXP (find_reg_note (last, REG_RETVAL, NULL_RTX), 0);
1116 insn != last; insn = NEXT_INSN (insn))
1117 {
1118 if (GET_CODE (insn) == CALL_INSN)
1119 benefit += 10; /* Assume at least this many insns in a library
1120 routine. */
1121 else if (GET_CODE (insn) == INSN
1122 && GET_CODE (PATTERN (insn)) != USE
1123 && GET_CODE (PATTERN (insn)) != CLOBBER)
1124 benefit++;
1125 }
1126
1127 return benefit;
1128 }
1129 \f
1130 /* Skip COUNT insns from INSN, counting library calls as 1 insn. */
1131
1132 static rtx
1133 skip_consec_insns (insn, count)
1134 rtx insn;
1135 int count;
1136 {
1137 for (; count > 0; count--)
1138 {
1139 rtx temp;
1140
1141 /* If first insn of libcall sequence, skip to end. */
1142 /* Do this at start of loop, since INSN is guaranteed to
1143 be an insn here. */
1144 if (GET_CODE (insn) != NOTE
1145 && (temp = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
1146 insn = XEXP (temp, 0);
1147
1148 do insn = NEXT_INSN (insn);
1149 while (GET_CODE (insn) == NOTE);
1150 }
1151
1152 return insn;
1153 }
1154
1155 /* Ignore any movable whose insn falls within a libcall
1156 which is part of another movable.
1157 We make use of the fact that the movable for the libcall value
1158 was made later and so appears later on the chain. */
1159
1160 static void
1161 ignore_some_movables (movables)
1162 struct movable *movables;
1163 {
1164 register struct movable *m, *m1;
1165
1166 for (m = movables; m; m = m->next)
1167 {
1168 /* Is this a movable for the value of a libcall? */
1169 rtx note = find_reg_note (m->insn, REG_RETVAL, NULL_RTX);
1170 if (note)
1171 {
1172 rtx insn;
1173 /* Check for earlier movables inside that range,
1174 and mark them invalid. We cannot use LUIDs here because
1175 insns created by loop.c for prior loops don't have LUIDs.
1176 Rather than reject all such insns from movables, we just
1177 explicitly check each insn in the libcall (since invariant
1178 libcalls aren't that common). */
1179 for (insn = XEXP (note, 0); insn != m->insn; insn = NEXT_INSN (insn))
1180 for (m1 = movables; m1 != m; m1 = m1->next)
1181 if (m1->insn == insn)
1182 m1->done = 1;
1183 }
1184 }
1185 }
1186
1187 /* For each movable insn, see if the reg that it loads
1188 leads when it dies right into another conditionally movable insn.
1189 If so, record that the second insn "forces" the first one,
1190 since the second can be moved only if the first is. */
1191
1192 static void
1193 force_movables (movables)
1194 struct movable *movables;
1195 {
1196 register struct movable *m, *m1;
1197 for (m1 = movables; m1; m1 = m1->next)
1198 /* Omit this if moving just the (SET (REG) 0) of a zero-extend. */
1199 if (!m1->partial && !m1->done)
1200 {
1201 int regno = m1->regno;
1202 for (m = m1->next; m; m = m->next)
1203 /* ??? Could this be a bug? What if CSE caused the
1204 register of M1 to be used after this insn?
1205 Since CSE does not update regno_last_uid,
1206 this insn M->insn might not be where it dies.
1207 But very likely this doesn't matter; what matters is
1208 that M's reg is computed from M1's reg. */
1209 if (INSN_UID (m->insn) == regno_last_uid[regno]
1210 && !m->done)
1211 break;
1212 if (m != 0 && m->set_src == m1->set_dest
1213 /* If m->consec, m->set_src isn't valid. */
1214 && m->consec == 0)
1215 m = 0;
1216
1217 /* Increase the priority of the moving the first insn
1218 since it permits the second to be moved as well. */
1219 if (m != 0)
1220 {
1221 m->forces = m1;
1222 m1->lifetime += m->lifetime;
1223 m1->savings += m1->savings;
1224 }
1225 }
1226 }
1227 \f
1228 /* Find invariant expressions that are equal and can be combined into
1229 one register. */
1230
1231 static void
1232 combine_movables (movables, nregs)
1233 struct movable *movables;
1234 int nregs;
1235 {
1236 register struct movable *m;
1237 char *matched_regs = (char *) alloca (nregs);
1238 enum machine_mode mode;
1239
1240 /* Regs that are set more than once are not allowed to match
1241 or be matched. I'm no longer sure why not. */
1242 /* Perhaps testing m->consec_sets would be more appropriate here? */
1243
1244 for (m = movables; m; m = m->next)
1245 if (m->match == 0 && n_times_used[m->regno] == 1 && !m->partial)
1246 {
1247 register struct movable *m1;
1248 int regno = m->regno;
1249
1250 bzero (matched_regs, nregs);
1251 matched_regs[regno] = 1;
1252
1253 for (m1 = movables; m1; m1 = m1->next)
1254 if (m != m1 && m1->match == 0 && n_times_used[m1->regno] == 1
1255 /* A reg used outside the loop mustn't be eliminated. */
1256 && !m1->global
1257 /* A reg used for zero-extending mustn't be eliminated. */
1258 && !m1->partial
1259 && (matched_regs[m1->regno]
1260 ||
1261 (
1262 /* Can combine regs with different modes loaded from the
1263 same constant only if the modes are the same or
1264 if both are integer modes with M wider or the same
1265 width as M1. The check for integer is redundant, but
1266 safe, since the only case of differing destination
1267 modes with equal sources is when both sources are
1268 VOIDmode, i.e., CONST_INT. */
1269 (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest)
1270 || (GET_MODE_CLASS (GET_MODE (m->set_dest)) == MODE_INT
1271 && GET_MODE_CLASS (GET_MODE (m1->set_dest)) == MODE_INT
1272 && (GET_MODE_BITSIZE (GET_MODE (m->set_dest))
1273 >= GET_MODE_BITSIZE (GET_MODE (m1->set_dest)))))
1274 /* See if the source of M1 says it matches M. */
1275 && ((GET_CODE (m1->set_src) == REG
1276 && matched_regs[REGNO (m1->set_src)])
1277 || rtx_equal_for_loop_p (m->set_src, m1->set_src,
1278 movables))))
1279 && ((m->dependencies == m1->dependencies)
1280 || rtx_equal_p (m->dependencies, m1->dependencies)))
1281 {
1282 m->lifetime += m1->lifetime;
1283 m->savings += m1->savings;
1284 m1->done = 1;
1285 m1->match = m;
1286 matched_regs[m1->regno] = 1;
1287 }
1288 }
1289
1290 /* Now combine the regs used for zero-extension.
1291 This can be done for those not marked `global'
1292 provided their lives don't overlap. */
1293
1294 for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode;
1295 mode = GET_MODE_WIDER_MODE (mode))
1296 {
1297 register struct movable *m0 = 0;
1298
1299 /* Combine all the registers for extension from mode MODE.
1300 Don't combine any that are used outside this loop. */
1301 for (m = movables; m; m = m->next)
1302 if (m->partial && ! m->global
1303 && mode == GET_MODE (SET_SRC (PATTERN (NEXT_INSN (m->insn)))))
1304 {
1305 register struct movable *m1;
1306 int first = uid_luid[regno_first_uid[m->regno]];
1307 int last = uid_luid[regno_last_uid[m->regno]];
1308
1309 if (m0 == 0)
1310 {
1311 /* First one: don't check for overlap, just record it. */
1312 m0 = m;
1313 continue;
1314 }
1315
1316 /* Make sure they extend to the same mode.
1317 (Almost always true.) */
1318 if (GET_MODE (m->set_dest) != GET_MODE (m0->set_dest))
1319 continue;
1320
1321 /* We already have one: check for overlap with those
1322 already combined together. */
1323 for (m1 = movables; m1 != m; m1 = m1->next)
1324 if (m1 == m0 || (m1->partial && m1->match == m0))
1325 if (! (uid_luid[regno_first_uid[m1->regno]] > last
1326 || uid_luid[regno_last_uid[m1->regno]] < first))
1327 goto overlap;
1328
1329 /* No overlap: we can combine this with the others. */
1330 m0->lifetime += m->lifetime;
1331 m0->savings += m->savings;
1332 m->done = 1;
1333 m->match = m0;
1334
1335 overlap: ;
1336 }
1337 }
1338 }
1339 \f
1340 /* Return 1 if regs X and Y will become the same if moved. */
1341
1342 static int
1343 regs_match_p (x, y, movables)
1344 rtx x, y;
1345 struct movable *movables;
1346 {
1347 int xn = REGNO (x);
1348 int yn = REGNO (y);
1349 struct movable *mx, *my;
1350
1351 for (mx = movables; mx; mx = mx->next)
1352 if (mx->regno == xn)
1353 break;
1354
1355 for (my = movables; my; my = my->next)
1356 if (my->regno == yn)
1357 break;
1358
1359 return (mx && my
1360 && ((mx->match == my->match && mx->match != 0)
1361 || mx->match == my
1362 || mx == my->match));
1363 }
1364
1365 /* Return 1 if X and Y are identical-looking rtx's.
1366 This is the Lisp function EQUAL for rtx arguments.
1367
1368 If two registers are matching movables or a movable register and an
1369 equivalent constant, consider them equal. */
1370
1371 static int
1372 rtx_equal_for_loop_p (x, y, movables)
1373 rtx x, y;
1374 struct movable *movables;
1375 {
1376 register int i;
1377 register int j;
1378 register struct movable *m;
1379 register enum rtx_code code;
1380 register char *fmt;
1381
1382 if (x == y)
1383 return 1;
1384 if (x == 0 || y == 0)
1385 return 0;
1386
1387 code = GET_CODE (x);
1388
1389 /* If we have a register and a constant, they may sometimes be
1390 equal. */
1391 if (GET_CODE (x) == REG && n_times_set[REGNO (x)] == -2
1392 && CONSTANT_P (y))
1393 for (m = movables; m; m = m->next)
1394 if (m->move_insn && m->regno == REGNO (x)
1395 && rtx_equal_p (m->set_src, y))
1396 return 1;
1397
1398 else if (GET_CODE (y) == REG && n_times_set[REGNO (y)] == -2
1399 && CONSTANT_P (x))
1400 for (m = movables; m; m = m->next)
1401 if (m->move_insn && m->regno == REGNO (y)
1402 && rtx_equal_p (m->set_src, x))
1403 return 1;
1404
1405 /* Otherwise, rtx's of different codes cannot be equal. */
1406 if (code != GET_CODE (y))
1407 return 0;
1408
1409 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1410 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1411
1412 if (GET_MODE (x) != GET_MODE (y))
1413 return 0;
1414
1415 /* These three types of rtx's can be compared nonrecursively. */
1416 if (code == REG)
1417 return (REGNO (x) == REGNO (y) || regs_match_p (x, y, movables));
1418
1419 if (code == LABEL_REF)
1420 return XEXP (x, 0) == XEXP (y, 0);
1421 if (code == SYMBOL_REF)
1422 return XSTR (x, 0) == XSTR (y, 0);
1423
1424 /* Compare the elements. If any pair of corresponding elements
1425 fail to match, return 0 for the whole things. */
1426
1427 fmt = GET_RTX_FORMAT (code);
1428 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1429 {
1430 switch (fmt[i])
1431 {
1432 case 'w':
1433 if (XWINT (x, i) != XWINT (y, i))
1434 return 0;
1435 break;
1436
1437 case 'i':
1438 if (XINT (x, i) != XINT (y, i))
1439 return 0;
1440 break;
1441
1442 case 'E':
1443 /* Two vectors must have the same length. */
1444 if (XVECLEN (x, i) != XVECLEN (y, i))
1445 return 0;
1446
1447 /* And the corresponding elements must match. */
1448 for (j = 0; j < XVECLEN (x, i); j++)
1449 if (rtx_equal_for_loop_p (XVECEXP (x, i, j), XVECEXP (y, i, j), movables) == 0)
1450 return 0;
1451 break;
1452
1453 case 'e':
1454 if (rtx_equal_for_loop_p (XEXP (x, i), XEXP (y, i), movables) == 0)
1455 return 0;
1456 break;
1457
1458 case 's':
1459 if (strcmp (XSTR (x, i), XSTR (y, i)))
1460 return 0;
1461 break;
1462
1463 case 'u':
1464 /* These are just backpointers, so they don't matter. */
1465 break;
1466
1467 case '0':
1468 break;
1469
1470 /* It is believed that rtx's at this level will never
1471 contain anything but integers and other rtx's,
1472 except for within LABEL_REFs and SYMBOL_REFs. */
1473 default:
1474 abort ();
1475 }
1476 }
1477 return 1;
1478 }
1479 \f
1480 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to all
1481 insns in INSNS which use thet reference. */
1482
1483 static void
1484 add_label_notes (x, insns)
1485 rtx x;
1486 rtx insns;
1487 {
1488 enum rtx_code code = GET_CODE (x);
1489 int i, j;
1490 char *fmt;
1491 rtx insn;
1492
1493 if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
1494 {
1495 rtx next = next_real_insn (XEXP (x, 0));
1496
1497 /* Don't record labels that refer to dispatch tables.
1498 This is not necessary, since the tablejump references the same label.
1499 And if we did record them, flow.c would make worse code. */
1500 if (next == 0
1501 || ! (GET_CODE (next) == JUMP_INSN
1502 && (GET_CODE (PATTERN (next)) == ADDR_VEC
1503 || GET_CODE (PATTERN (next)) == ADDR_DIFF_VEC)))
1504 {
1505 for (insn = insns; insn; insn = NEXT_INSN (insn))
1506 if (reg_mentioned_p (XEXP (x, 0), insn))
1507 REG_NOTES (insn) = gen_rtx (EXPR_LIST, REG_LABEL, XEXP (x, 0),
1508 REG_NOTES (insn));
1509 }
1510 return;
1511 }
1512
1513 fmt = GET_RTX_FORMAT (code);
1514 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1515 {
1516 if (fmt[i] == 'e')
1517 add_label_notes (XEXP (x, i), insns);
1518 else if (fmt[i] == 'E')
1519 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1520 add_label_notes (XVECEXP (x, i, j), insns);
1521 }
1522 }
1523 \f
1524 /* Scan MOVABLES, and move the insns that deserve to be moved.
1525 If two matching movables are combined, replace one reg with the
1526 other throughout. */
1527
1528 static void
1529 move_movables (movables, threshold, insn_count, loop_start, end, nregs)
1530 struct movable *movables;
1531 int threshold;
1532 int insn_count;
1533 rtx loop_start;
1534 rtx end;
1535 int nregs;
1536 {
1537 rtx new_start = 0;
1538 register struct movable *m;
1539 register rtx p;
1540 /* Map of pseudo-register replacements to handle combining
1541 when we move several insns that load the same value
1542 into different pseudo-registers. */
1543 rtx *reg_map = (rtx *) alloca (nregs * sizeof (rtx));
1544 char *already_moved = (char *) alloca (nregs);
1545
1546 bzero (already_moved, nregs);
1547 bzero ((char *) reg_map, nregs * sizeof (rtx));
1548
1549 num_movables = 0;
1550
1551 for (m = movables; m; m = m->next)
1552 {
1553 /* Describe this movable insn. */
1554
1555 if (loop_dump_stream)
1556 {
1557 fprintf (loop_dump_stream, "Insn %d: regno %d (life %d), ",
1558 INSN_UID (m->insn), m->regno, m->lifetime);
1559 if (m->consec > 0)
1560 fprintf (loop_dump_stream, "consec %d, ", m->consec);
1561 if (m->cond)
1562 fprintf (loop_dump_stream, "cond ");
1563 if (m->force)
1564 fprintf (loop_dump_stream, "force ");
1565 if (m->global)
1566 fprintf (loop_dump_stream, "global ");
1567 if (m->done)
1568 fprintf (loop_dump_stream, "done ");
1569 if (m->move_insn)
1570 fprintf (loop_dump_stream, "move-insn ");
1571 if (m->match)
1572 fprintf (loop_dump_stream, "matches %d ",
1573 INSN_UID (m->match->insn));
1574 if (m->forces)
1575 fprintf (loop_dump_stream, "forces %d ",
1576 INSN_UID (m->forces->insn));
1577 }
1578
1579 /* Count movables. Value used in heuristics in strength_reduce. */
1580 num_movables++;
1581
1582 /* Ignore the insn if it's already done (it matched something else).
1583 Otherwise, see if it is now safe to move. */
1584
1585 if (!m->done
1586 && (! m->cond
1587 || (1 == invariant_p (m->set_src)
1588 && (m->dependencies == 0
1589 || 1 == invariant_p (m->dependencies))
1590 && (m->consec == 0
1591 || 1 == consec_sets_invariant_p (m->set_dest,
1592 m->consec + 1,
1593 m->insn))))
1594 && (! m->forces || m->forces->done))
1595 {
1596 register int regno;
1597 register rtx p;
1598 int savings = m->savings;
1599
1600 /* We have an insn that is safe to move.
1601 Compute its desirability. */
1602
1603 p = m->insn;
1604 regno = m->regno;
1605
1606 if (loop_dump_stream)
1607 fprintf (loop_dump_stream, "savings %d ", savings);
1608
1609 if (moved_once[regno])
1610 {
1611 insn_count *= 2;
1612
1613 if (loop_dump_stream)
1614 fprintf (loop_dump_stream, "halved since already moved ");
1615 }
1616
1617 /* An insn MUST be moved if we already moved something else
1618 which is safe only if this one is moved too: that is,
1619 if already_moved[REGNO] is nonzero. */
1620
1621 /* An insn is desirable to move if the new lifetime of the
1622 register is no more than THRESHOLD times the old lifetime.
1623 If it's not desirable, it means the loop is so big
1624 that moving won't speed things up much,
1625 and it is liable to make register usage worse. */
1626
1627 /* It is also desirable to move if it can be moved at no
1628 extra cost because something else was already moved. */
1629
1630 if (already_moved[regno]
1631 || (threshold * savings * m->lifetime) >= insn_count
1632 || (m->forces && m->forces->done
1633 && n_times_used[m->forces->regno] == 1))
1634 {
1635 int count;
1636 register struct movable *m1;
1637 rtx first;
1638
1639 /* Now move the insns that set the reg. */
1640
1641 if (m->partial && m->match)
1642 {
1643 rtx newpat, i1;
1644 rtx r1, r2;
1645 /* Find the end of this chain of matching regs.
1646 Thus, we load each reg in the chain from that one reg.
1647 And that reg is loaded with 0 directly,
1648 since it has ->match == 0. */
1649 for (m1 = m; m1->match; m1 = m1->match);
1650 newpat = gen_move_insn (SET_DEST (PATTERN (m->insn)),
1651 SET_DEST (PATTERN (m1->insn)));
1652 i1 = emit_insn_before (newpat, loop_start);
1653
1654 /* Mark the moved, invariant reg as being allowed to
1655 share a hard reg with the other matching invariant. */
1656 REG_NOTES (i1) = REG_NOTES (m->insn);
1657 r1 = SET_DEST (PATTERN (m->insn));
1658 r2 = SET_DEST (PATTERN (m1->insn));
1659 regs_may_share = gen_rtx (EXPR_LIST, VOIDmode, r1,
1660 gen_rtx (EXPR_LIST, VOIDmode, r2,
1661 regs_may_share));
1662 delete_insn (m->insn);
1663
1664 if (new_start == 0)
1665 new_start = i1;
1666
1667 if (loop_dump_stream)
1668 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1669 }
1670 /* If we are to re-generate the item being moved with a
1671 new move insn, first delete what we have and then emit
1672 the move insn before the loop. */
1673 else if (m->move_insn)
1674 {
1675 rtx i1, temp;
1676
1677 for (count = m->consec; count >= 0; count--)
1678 {
1679 /* If this is the first insn of a library call sequence,
1680 skip to the end. */
1681 if (GET_CODE (p) != NOTE
1682 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1683 p = XEXP (temp, 0);
1684
1685 /* If this is the last insn of a libcall sequence, then
1686 delete every insn in the sequence except the last.
1687 The last insn is handled in the normal manner. */
1688 if (GET_CODE (p) != NOTE
1689 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1690 {
1691 temp = XEXP (temp, 0);
1692 while (temp != p)
1693 temp = delete_insn (temp);
1694 }
1695
1696 p = delete_insn (p);
1697 while (p && GET_CODE (p) == NOTE)
1698 p = NEXT_INSN (p);
1699 }
1700
1701 start_sequence ();
1702 emit_move_insn (m->set_dest, m->set_src);
1703 temp = get_insns ();
1704 end_sequence ();
1705
1706 add_label_notes (m->set_src, temp);
1707
1708 i1 = emit_insns_before (temp, loop_start);
1709 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
1710 REG_NOTES (i1)
1711 = gen_rtx (EXPR_LIST,
1712 m->is_equiv ? REG_EQUIV : REG_EQUAL,
1713 m->set_src, REG_NOTES (i1));
1714
1715 if (loop_dump_stream)
1716 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1717
1718 /* The more regs we move, the less we like moving them. */
1719 threshold -= 3;
1720 }
1721 else
1722 {
1723 for (count = m->consec; count >= 0; count--)
1724 {
1725 rtx i1, temp;
1726
1727 /* If first insn of libcall sequence, skip to end. */
1728 /* Do this at start of loop, since p is guaranteed to
1729 be an insn here. */
1730 if (GET_CODE (p) != NOTE
1731 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1732 p = XEXP (temp, 0);
1733
1734 /* If last insn of libcall sequence, move all
1735 insns except the last before the loop. The last
1736 insn is handled in the normal manner. */
1737 if (GET_CODE (p) != NOTE
1738 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1739 {
1740 rtx fn_address = 0;
1741 rtx fn_reg = 0;
1742 rtx fn_address_insn = 0;
1743
1744 first = 0;
1745 for (temp = XEXP (temp, 0); temp != p;
1746 temp = NEXT_INSN (temp))
1747 {
1748 rtx body;
1749 rtx n;
1750 rtx next;
1751
1752 if (GET_CODE (temp) == NOTE)
1753 continue;
1754
1755 body = PATTERN (temp);
1756
1757 /* Find the next insn after TEMP,
1758 not counting USE or NOTE insns. */
1759 for (next = NEXT_INSN (temp); next != p;
1760 next = NEXT_INSN (next))
1761 if (! (GET_CODE (next) == INSN
1762 && GET_CODE (PATTERN (next)) == USE)
1763 && GET_CODE (next) != NOTE)
1764 break;
1765
1766 /* If that is the call, this may be the insn
1767 that loads the function address.
1768
1769 Extract the function address from the insn
1770 that loads it into a register.
1771 If this insn was cse'd, we get incorrect code.
1772
1773 So emit a new move insn that copies the
1774 function address into the register that the
1775 call insn will use. flow.c will delete any
1776 redundant stores that we have created. */
1777 if (GET_CODE (next) == CALL_INSN
1778 && GET_CODE (body) == SET
1779 && GET_CODE (SET_DEST (body)) == REG
1780 && (n = find_reg_note (temp, REG_EQUAL,
1781 NULL_RTX)))
1782 {
1783 fn_reg = SET_SRC (body);
1784 if (GET_CODE (fn_reg) != REG)
1785 fn_reg = SET_DEST (body);
1786 fn_address = XEXP (n, 0);
1787 fn_address_insn = temp;
1788 }
1789 /* We have the call insn.
1790 If it uses the register we suspect it might,
1791 load it with the correct address directly. */
1792 if (GET_CODE (temp) == CALL_INSN
1793 && fn_address != 0
1794 && reg_referenced_p (fn_reg, body))
1795 emit_insn_after (gen_move_insn (fn_reg,
1796 fn_address),
1797 fn_address_insn);
1798
1799 if (GET_CODE (temp) == CALL_INSN)
1800 {
1801 i1 = emit_call_insn_before (body, loop_start);
1802 /* Because the USAGE information potentially
1803 contains objects other than hard registers
1804 we need to copy it. */
1805 if (CALL_INSN_FUNCTION_USAGE (temp))
1806 CALL_INSN_FUNCTION_USAGE (i1) =
1807 copy_rtx (CALL_INSN_FUNCTION_USAGE (temp));
1808 }
1809 else
1810 i1 = emit_insn_before (body, loop_start);
1811 if (first == 0)
1812 first = i1;
1813 if (temp == fn_address_insn)
1814 fn_address_insn = i1;
1815 REG_NOTES (i1) = REG_NOTES (temp);
1816 delete_insn (temp);
1817 }
1818 }
1819 if (m->savemode != VOIDmode)
1820 {
1821 /* P sets REG to zero; but we should clear only
1822 the bits that are not covered by the mode
1823 m->savemode. */
1824 rtx reg = m->set_dest;
1825 rtx sequence;
1826 rtx tem;
1827
1828 start_sequence ();
1829 tem = expand_binop
1830 (GET_MODE (reg), and_optab, reg,
1831 GEN_INT ((((HOST_WIDE_INT) 1
1832 << GET_MODE_BITSIZE (m->savemode)))
1833 - 1),
1834 reg, 1, OPTAB_LIB_WIDEN);
1835 if (tem == 0)
1836 abort ();
1837 if (tem != reg)
1838 emit_move_insn (reg, tem);
1839 sequence = gen_sequence ();
1840 end_sequence ();
1841 i1 = emit_insn_before (sequence, loop_start);
1842 }
1843 else if (GET_CODE (p) == CALL_INSN)
1844 {
1845 i1 = emit_call_insn_before (PATTERN (p), loop_start);
1846 /* Because the USAGE information potentially
1847 contains objects other than hard registers
1848 we need to copy it. */
1849 if (CALL_INSN_FUNCTION_USAGE (p))
1850 CALL_INSN_FUNCTION_USAGE (i1) =
1851 copy_rtx (CALL_INSN_FUNCTION_USAGE (p));
1852 }
1853 else
1854 i1 = emit_insn_before (PATTERN (p), loop_start);
1855
1856 REG_NOTES (i1) = REG_NOTES (p);
1857
1858 /* If there is a REG_EQUAL note present whose value is
1859 not loop invariant, then delete it, since it may
1860 cause problems with later optimization passes.
1861 It is possible for cse to create such notes
1862 like this as a result of record_jump_cond. */
1863
1864 if ((temp = find_reg_note (i1, REG_EQUAL, NULL_RTX))
1865 && ! invariant_p (XEXP (temp, 0)))
1866 remove_note (i1, temp);
1867
1868 if (new_start == 0)
1869 new_start = i1;
1870
1871 if (loop_dump_stream)
1872 fprintf (loop_dump_stream, " moved to %d",
1873 INSN_UID (i1));
1874
1875 #if 0
1876 /* This isn't needed because REG_NOTES is copied
1877 below and is wrong since P might be a PARALLEL. */
1878 if (REG_NOTES (i1) == 0
1879 && ! m->partial /* But not if it's a zero-extend clr. */
1880 && ! m->global /* and not if used outside the loop
1881 (since it might get set outside). */
1882 && CONSTANT_P (SET_SRC (PATTERN (p))))
1883 REG_NOTES (i1)
1884 = gen_rtx (EXPR_LIST, REG_EQUAL,
1885 SET_SRC (PATTERN (p)), REG_NOTES (i1));
1886 #endif
1887
1888 /* If library call, now fix the REG_NOTES that contain
1889 insn pointers, namely REG_LIBCALL on FIRST
1890 and REG_RETVAL on I1. */
1891 if (temp = find_reg_note (i1, REG_RETVAL, NULL_RTX))
1892 {
1893 XEXP (temp, 0) = first;
1894 temp = find_reg_note (first, REG_LIBCALL, NULL_RTX);
1895 XEXP (temp, 0) = i1;
1896 }
1897
1898 delete_insn (p);
1899 do p = NEXT_INSN (p);
1900 while (p && GET_CODE (p) == NOTE);
1901 }
1902
1903 /* The more regs we move, the less we like moving them. */
1904 threshold -= 3;
1905 }
1906
1907 /* Any other movable that loads the same register
1908 MUST be moved. */
1909 already_moved[regno] = 1;
1910
1911 /* This reg has been moved out of one loop. */
1912 moved_once[regno] = 1;
1913
1914 /* The reg set here is now invariant. */
1915 if (! m->partial)
1916 n_times_set[regno] = 0;
1917
1918 m->done = 1;
1919
1920 /* Change the length-of-life info for the register
1921 to say it lives at least the full length of this loop.
1922 This will help guide optimizations in outer loops. */
1923
1924 if (uid_luid[regno_first_uid[regno]] > INSN_LUID (loop_start))
1925 /* This is the old insn before all the moved insns.
1926 We can't use the moved insn because it is out of range
1927 in uid_luid. Only the old insns have luids. */
1928 regno_first_uid[regno] = INSN_UID (loop_start);
1929 if (uid_luid[regno_last_uid[regno]] < INSN_LUID (end))
1930 regno_last_uid[regno] = INSN_UID (end);
1931
1932 /* Combine with this moved insn any other matching movables. */
1933
1934 if (! m->partial)
1935 for (m1 = movables; m1; m1 = m1->next)
1936 if (m1->match == m)
1937 {
1938 rtx temp;
1939
1940 /* Schedule the reg loaded by M1
1941 for replacement so that shares the reg of M.
1942 If the modes differ (only possible in restricted
1943 circumstances, make a SUBREG. */
1944 if (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest))
1945 reg_map[m1->regno] = m->set_dest;
1946 else
1947 reg_map[m1->regno]
1948 = gen_lowpart_common (GET_MODE (m1->set_dest),
1949 m->set_dest);
1950
1951 /* Get rid of the matching insn
1952 and prevent further processing of it. */
1953 m1->done = 1;
1954
1955 /* if library call, delete all insn except last, which
1956 is deleted below */
1957 if (temp = find_reg_note (m1->insn, REG_RETVAL,
1958 NULL_RTX))
1959 {
1960 for (temp = XEXP (temp, 0); temp != m1->insn;
1961 temp = NEXT_INSN (temp))
1962 delete_insn (temp);
1963 }
1964 delete_insn (m1->insn);
1965
1966 /* Any other movable that loads the same register
1967 MUST be moved. */
1968 already_moved[m1->regno] = 1;
1969
1970 /* The reg merged here is now invariant,
1971 if the reg it matches is invariant. */
1972 if (! m->partial)
1973 n_times_set[m1->regno] = 0;
1974 }
1975 }
1976 else if (loop_dump_stream)
1977 fprintf (loop_dump_stream, "not desirable");
1978 }
1979 else if (loop_dump_stream && !m->match)
1980 fprintf (loop_dump_stream, "not safe");
1981
1982 if (loop_dump_stream)
1983 fprintf (loop_dump_stream, "\n");
1984 }
1985
1986 if (new_start == 0)
1987 new_start = loop_start;
1988
1989 /* Go through all the instructions in the loop, making
1990 all the register substitutions scheduled in REG_MAP. */
1991 for (p = new_start; p != end; p = NEXT_INSN (p))
1992 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1993 || GET_CODE (p) == CALL_INSN)
1994 {
1995 replace_regs (PATTERN (p), reg_map, nregs, 0);
1996 replace_regs (REG_NOTES (p), reg_map, nregs, 0);
1997 INSN_CODE (p) = -1;
1998 }
1999 }
2000 \f
2001 #if 0
2002 /* Scan X and replace the address of any MEM in it with ADDR.
2003 REG is the address that MEM should have before the replacement. */
2004
2005 static void
2006 replace_call_address (x, reg, addr)
2007 rtx x, reg, addr;
2008 {
2009 register enum rtx_code code;
2010 register int i;
2011 register char *fmt;
2012
2013 if (x == 0)
2014 return;
2015 code = GET_CODE (x);
2016 switch (code)
2017 {
2018 case PC:
2019 case CC0:
2020 case CONST_INT:
2021 case CONST_DOUBLE:
2022 case CONST:
2023 case SYMBOL_REF:
2024 case LABEL_REF:
2025 case REG:
2026 return;
2027
2028 case SET:
2029 /* Short cut for very common case. */
2030 replace_call_address (XEXP (x, 1), reg, addr);
2031 return;
2032
2033 case CALL:
2034 /* Short cut for very common case. */
2035 replace_call_address (XEXP (x, 0), reg, addr);
2036 return;
2037
2038 case MEM:
2039 /* If this MEM uses a reg other than the one we expected,
2040 something is wrong. */
2041 if (XEXP (x, 0) != reg)
2042 abort ();
2043 XEXP (x, 0) = addr;
2044 return;
2045 }
2046
2047 fmt = GET_RTX_FORMAT (code);
2048 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2049 {
2050 if (fmt[i] == 'e')
2051 replace_call_address (XEXP (x, i), reg, addr);
2052 if (fmt[i] == 'E')
2053 {
2054 register int j;
2055 for (j = 0; j < XVECLEN (x, i); j++)
2056 replace_call_address (XVECEXP (x, i, j), reg, addr);
2057 }
2058 }
2059 }
2060 #endif
2061 \f
2062 /* Return the number of memory refs to addresses that vary
2063 in the rtx X. */
2064
2065 static int
2066 count_nonfixed_reads (x)
2067 rtx x;
2068 {
2069 register enum rtx_code code;
2070 register int i;
2071 register char *fmt;
2072 int value;
2073
2074 if (x == 0)
2075 return 0;
2076
2077 code = GET_CODE (x);
2078 switch (code)
2079 {
2080 case PC:
2081 case CC0:
2082 case CONST_INT:
2083 case CONST_DOUBLE:
2084 case CONST:
2085 case SYMBOL_REF:
2086 case LABEL_REF:
2087 case REG:
2088 return 0;
2089
2090 case MEM:
2091 return ((invariant_p (XEXP (x, 0)) != 1)
2092 + count_nonfixed_reads (XEXP (x, 0)));
2093 }
2094
2095 value = 0;
2096 fmt = GET_RTX_FORMAT (code);
2097 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2098 {
2099 if (fmt[i] == 'e')
2100 value += count_nonfixed_reads (XEXP (x, i));
2101 if (fmt[i] == 'E')
2102 {
2103 register int j;
2104 for (j = 0; j < XVECLEN (x, i); j++)
2105 value += count_nonfixed_reads (XVECEXP (x, i, j));
2106 }
2107 }
2108 return value;
2109 }
2110
2111 \f
2112 #if 0
2113 /* P is an instruction that sets a register to the result of a ZERO_EXTEND.
2114 Replace it with an instruction to load just the low bytes
2115 if the machine supports such an instruction,
2116 and insert above LOOP_START an instruction to clear the register. */
2117
2118 static void
2119 constant_high_bytes (p, loop_start)
2120 rtx p, loop_start;
2121 {
2122 register rtx new;
2123 register int insn_code_number;
2124
2125 /* Try to change (SET (REG ...) (ZERO_EXTEND (..:B ...)))
2126 to (SET (STRICT_LOW_PART (SUBREG:B (REG...))) ...). */
2127
2128 new = gen_rtx (SET, VOIDmode,
2129 gen_rtx (STRICT_LOW_PART, VOIDmode,
2130 gen_rtx (SUBREG, GET_MODE (XEXP (SET_SRC (PATTERN (p)), 0)),
2131 SET_DEST (PATTERN (p)),
2132 0)),
2133 XEXP (SET_SRC (PATTERN (p)), 0));
2134 insn_code_number = recog (new, p);
2135
2136 if (insn_code_number)
2137 {
2138 register int i;
2139
2140 /* Clear destination register before the loop. */
2141 emit_insn_before (gen_rtx (SET, VOIDmode,
2142 SET_DEST (PATTERN (p)),
2143 const0_rtx),
2144 loop_start);
2145
2146 /* Inside the loop, just load the low part. */
2147 PATTERN (p) = new;
2148 }
2149 }
2150 #endif
2151 \f
2152 /* Scan a loop setting the variables `unknown_address_altered',
2153 `num_mem_sets', `loop_continue', loops_enclosed', `loop_has_call',
2154 and `loop_has_volatile'.
2155 Also, fill in the array `loop_store_mems'. */
2156
2157 static void
2158 prescan_loop (start, end)
2159 rtx start, end;
2160 {
2161 register int level = 1;
2162 register rtx insn;
2163
2164 unknown_address_altered = 0;
2165 loop_has_call = 0;
2166 loop_has_volatile = 0;
2167 loop_store_mems_idx = 0;
2168
2169 num_mem_sets = 0;
2170 loops_enclosed = 1;
2171 loop_continue = 0;
2172
2173 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2174 insn = NEXT_INSN (insn))
2175 {
2176 if (GET_CODE (insn) == NOTE)
2177 {
2178 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
2179 {
2180 ++level;
2181 /* Count number of loops contained in this one. */
2182 loops_enclosed++;
2183 }
2184 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END)
2185 {
2186 --level;
2187 if (level == 0)
2188 {
2189 end = insn;
2190 break;
2191 }
2192 }
2193 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_CONT)
2194 {
2195 if (level == 1)
2196 loop_continue = insn;
2197 }
2198 }
2199 else if (GET_CODE (insn) == CALL_INSN)
2200 {
2201 unknown_address_altered = 1;
2202 loop_has_call = 1;
2203 }
2204 else
2205 {
2206 if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
2207 {
2208 if (volatile_refs_p (PATTERN (insn)))
2209 loop_has_volatile = 1;
2210
2211 note_stores (PATTERN (insn), note_addr_stored);
2212 }
2213 }
2214 }
2215 }
2216 \f
2217 /* Scan the function looking for loops. Record the start and end of each loop.
2218 Also mark as invalid loops any loops that contain a setjmp or are branched
2219 to from outside the loop. */
2220
2221 static void
2222 find_and_verify_loops (f)
2223 rtx f;
2224 {
2225 rtx insn, label;
2226 int current_loop = -1;
2227 int next_loop = -1;
2228 int loop;
2229
2230 /* If there are jumps to undefined labels,
2231 treat them as jumps out of any/all loops.
2232 This also avoids writing past end of tables when there are no loops. */
2233 uid_loop_num[0] = -1;
2234
2235 /* Find boundaries of loops, mark which loops are contained within
2236 loops, and invalidate loops that have setjmp. */
2237
2238 for (insn = f; insn; insn = NEXT_INSN (insn))
2239 {
2240 if (GET_CODE (insn) == NOTE)
2241 switch (NOTE_LINE_NUMBER (insn))
2242 {
2243 case NOTE_INSN_LOOP_BEG:
2244 loop_number_loop_starts[++next_loop] = insn;
2245 loop_number_loop_ends[next_loop] = 0;
2246 loop_outer_loop[next_loop] = current_loop;
2247 loop_invalid[next_loop] = 0;
2248 loop_number_exit_labels[next_loop] = 0;
2249 loop_number_exit_count[next_loop] = 0;
2250 current_loop = next_loop;
2251 break;
2252
2253 case NOTE_INSN_SETJMP:
2254 /* In this case, we must invalidate our current loop and any
2255 enclosing loop. */
2256 for (loop = current_loop; loop != -1; loop = loop_outer_loop[loop])
2257 {
2258 loop_invalid[loop] = 1;
2259 if (loop_dump_stream)
2260 fprintf (loop_dump_stream,
2261 "\nLoop at %d ignored due to setjmp.\n",
2262 INSN_UID (loop_number_loop_starts[loop]));
2263 }
2264 break;
2265
2266 case NOTE_INSN_LOOP_END:
2267 if (current_loop == -1)
2268 abort ();
2269
2270 loop_number_loop_ends[current_loop] = insn;
2271 current_loop = loop_outer_loop[current_loop];
2272 break;
2273
2274 }
2275
2276 /* Note that this will mark the NOTE_INSN_LOOP_END note as being in the
2277 enclosing loop, but this doesn't matter. */
2278 uid_loop_num[INSN_UID (insn)] = current_loop;
2279 }
2280
2281 /* Any loop containing a label used in an initializer must be invalidated,
2282 because it can be jumped into from anywhere. */
2283
2284 for (label = forced_labels; label; label = XEXP (label, 1))
2285 {
2286 int loop_num;
2287
2288 for (loop_num = uid_loop_num[INSN_UID (XEXP (label, 0))];
2289 loop_num != -1;
2290 loop_num = loop_outer_loop[loop_num])
2291 loop_invalid[loop_num] = 1;
2292 }
2293
2294 /* Any loop containing a label used for an exception handler must be
2295 invalidated, because it can be jumped into from anywhere. */
2296
2297 for (label = exception_handler_labels; label; label = XEXP (label, 1))
2298 {
2299 int loop_num;
2300
2301 for (loop_num = uid_loop_num[INSN_UID (XEXP (label, 0))];
2302 loop_num != -1;
2303 loop_num = loop_outer_loop[loop_num])
2304 loop_invalid[loop_num] = 1;
2305 }
2306
2307 /* Now scan all insn's in the function. If any JUMP_INSN branches into a
2308 loop that it is not contained within, that loop is marked invalid.
2309 If any INSN or CALL_INSN uses a label's address, then the loop containing
2310 that label is marked invalid, because it could be jumped into from
2311 anywhere.
2312
2313 Also look for blocks of code ending in an unconditional branch that
2314 exits the loop. If such a block is surrounded by a conditional
2315 branch around the block, move the block elsewhere (see below) and
2316 invert the jump to point to the code block. This may eliminate a
2317 label in our loop and will simplify processing by both us and a
2318 possible second cse pass. */
2319
2320 for (insn = f; insn; insn = NEXT_INSN (insn))
2321 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
2322 {
2323 int this_loop_num = uid_loop_num[INSN_UID (insn)];
2324
2325 if (GET_CODE (insn) == INSN || GET_CODE (insn) == CALL_INSN)
2326 {
2327 rtx note = find_reg_note (insn, REG_LABEL, NULL_RTX);
2328 if (note)
2329 {
2330 int loop_num;
2331
2332 for (loop_num = uid_loop_num[INSN_UID (XEXP (note, 0))];
2333 loop_num != -1;
2334 loop_num = loop_outer_loop[loop_num])
2335 loop_invalid[loop_num] = 1;
2336 }
2337 }
2338
2339 if (GET_CODE (insn) != JUMP_INSN)
2340 continue;
2341
2342 mark_loop_jump (PATTERN (insn), this_loop_num);
2343
2344 /* See if this is an unconditional branch outside the loop. */
2345 if (this_loop_num != -1
2346 && (GET_CODE (PATTERN (insn)) == RETURN
2347 || (simplejump_p (insn)
2348 && (uid_loop_num[INSN_UID (JUMP_LABEL (insn))]
2349 != this_loop_num)))
2350 && get_max_uid () < max_uid_for_loop)
2351 {
2352 rtx p;
2353 rtx our_next = next_real_insn (insn);
2354 int dest_loop;
2355 int outer_loop = -1;
2356
2357 /* Go backwards until we reach the start of the loop, a label,
2358 or a JUMP_INSN. */
2359 for (p = PREV_INSN (insn);
2360 GET_CODE (p) != CODE_LABEL
2361 && ! (GET_CODE (p) == NOTE
2362 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
2363 && GET_CODE (p) != JUMP_INSN;
2364 p = PREV_INSN (p))
2365 ;
2366
2367 /* Check for the case where we have a jump to an inner nested
2368 loop, and do not perform the optimization in that case. */
2369
2370 if (JUMP_LABEL (insn))
2371 {
2372 dest_loop = uid_loop_num[INSN_UID (JUMP_LABEL (insn))];
2373 if (dest_loop != -1)
2374 {
2375 for (outer_loop = dest_loop; outer_loop != -1;
2376 outer_loop = loop_outer_loop[outer_loop])
2377 if (outer_loop == this_loop_num)
2378 break;
2379 }
2380 }
2381
2382 /* Make sure that the target of P is within the current loop. */
2383
2384 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
2385 && uid_loop_num[INSN_UID (JUMP_LABEL (p))] != this_loop_num)
2386 outer_loop = this_loop_num;
2387
2388 /* If we stopped on a JUMP_INSN to the next insn after INSN,
2389 we have a block of code to try to move.
2390
2391 We look backward and then forward from the target of INSN
2392 to find a BARRIER at the same loop depth as the target.
2393 If we find such a BARRIER, we make a new label for the start
2394 of the block, invert the jump in P and point it to that label,
2395 and move the block of code to the spot we found. */
2396
2397 if (outer_loop == -1
2398 && GET_CODE (p) == JUMP_INSN
2399 && JUMP_LABEL (p) != 0
2400 /* Just ignore jumps to labels that were never emitted.
2401 These always indicate compilation errors. */
2402 && INSN_UID (JUMP_LABEL (p)) != 0
2403 && condjump_p (p)
2404 && ! simplejump_p (p)
2405 && next_real_insn (JUMP_LABEL (p)) == our_next)
2406 {
2407 rtx target
2408 = JUMP_LABEL (insn) ? JUMP_LABEL (insn) : get_last_insn ();
2409 int target_loop_num = uid_loop_num[INSN_UID (target)];
2410 rtx loc;
2411
2412 for (loc = target; loc; loc = PREV_INSN (loc))
2413 if (GET_CODE (loc) == BARRIER
2414 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2415 break;
2416
2417 if (loc == 0)
2418 for (loc = target; loc; loc = NEXT_INSN (loc))
2419 if (GET_CODE (loc) == BARRIER
2420 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2421 break;
2422
2423 if (loc)
2424 {
2425 rtx cond_label = JUMP_LABEL (p);
2426 rtx new_label = get_label_after (p);
2427
2428 /* Ensure our label doesn't go away. */
2429 LABEL_NUSES (cond_label)++;
2430
2431 /* Verify that uid_loop_num is large enough and that
2432 we can invert P. */
2433 if (invert_jump (p, new_label))
2434 {
2435 rtx q, r;
2436
2437 /* Include the BARRIER after INSN and copy the
2438 block after LOC. */
2439 new_label = squeeze_notes (new_label, NEXT_INSN (insn));
2440 reorder_insns (new_label, NEXT_INSN (insn), loc);
2441
2442 /* All those insns are now in TARGET_LOOP_NUM. */
2443 for (q = new_label; q != NEXT_INSN (NEXT_INSN (insn));
2444 q = NEXT_INSN (q))
2445 uid_loop_num[INSN_UID (q)] = target_loop_num;
2446
2447 /* The label jumped to by INSN is no longer a loop exit.
2448 Unless INSN does not have a label (e.g., it is a
2449 RETURN insn), search loop_number_exit_labels to find
2450 its label_ref, and remove it. Also turn off
2451 LABEL_OUTSIDE_LOOP_P bit. */
2452 if (JUMP_LABEL (insn))
2453 {
2454 int loop_num;
2455
2456 for (q = 0,
2457 r = loop_number_exit_labels[this_loop_num];
2458 r; q = r, r = LABEL_NEXTREF (r))
2459 if (XEXP (r, 0) == JUMP_LABEL (insn))
2460 {
2461 LABEL_OUTSIDE_LOOP_P (r) = 0;
2462 if (q)
2463 LABEL_NEXTREF (q) = LABEL_NEXTREF (r);
2464 else
2465 loop_number_exit_labels[this_loop_num]
2466 = LABEL_NEXTREF (r);
2467 break;
2468 }
2469
2470 for (loop_num = this_loop_num;
2471 loop_num != -1 && loop_num != target_loop_num;
2472 loop_num = loop_outer_loop[loop_num])
2473 loop_number_exit_count[loop_num]--;
2474
2475 /* If we didn't find it, then something is wrong. */
2476 if (! r)
2477 abort ();
2478 }
2479
2480 /* P is now a jump outside the loop, so it must be put
2481 in loop_number_exit_labels, and marked as such.
2482 The easiest way to do this is to just call
2483 mark_loop_jump again for P. */
2484 mark_loop_jump (PATTERN (p), this_loop_num);
2485
2486 /* If INSN now jumps to the insn after it,
2487 delete INSN. */
2488 if (JUMP_LABEL (insn) != 0
2489 && (next_real_insn (JUMP_LABEL (insn))
2490 == next_real_insn (insn)))
2491 delete_insn (insn);
2492 }
2493
2494 /* Continue the loop after where the conditional
2495 branch used to jump, since the only branch insn
2496 in the block (if it still remains) is an inter-loop
2497 branch and hence needs no processing. */
2498 insn = NEXT_INSN (cond_label);
2499
2500 if (--LABEL_NUSES (cond_label) == 0)
2501 delete_insn (cond_label);
2502
2503 /* This loop will be continued with NEXT_INSN (insn). */
2504 insn = PREV_INSN (insn);
2505 }
2506 }
2507 }
2508 }
2509 }
2510
2511 /* If any label in X jumps to a loop different from LOOP_NUM and any of the
2512 loops it is contained in, mark the target loop invalid.
2513
2514 For speed, we assume that X is part of a pattern of a JUMP_INSN. */
2515
2516 static void
2517 mark_loop_jump (x, loop_num)
2518 rtx x;
2519 int loop_num;
2520 {
2521 int dest_loop;
2522 int outer_loop;
2523 int i;
2524
2525 switch (GET_CODE (x))
2526 {
2527 case PC:
2528 case USE:
2529 case CLOBBER:
2530 case REG:
2531 case MEM:
2532 case CONST_INT:
2533 case CONST_DOUBLE:
2534 case RETURN:
2535 return;
2536
2537 case CONST:
2538 /* There could be a label reference in here. */
2539 mark_loop_jump (XEXP (x, 0), loop_num);
2540 return;
2541
2542 case PLUS:
2543 case MINUS:
2544 case MULT:
2545 mark_loop_jump (XEXP (x, 0), loop_num);
2546 mark_loop_jump (XEXP (x, 1), loop_num);
2547 return;
2548
2549 case SIGN_EXTEND:
2550 case ZERO_EXTEND:
2551 mark_loop_jump (XEXP (x, 0), loop_num);
2552 return;
2553
2554 case LABEL_REF:
2555 dest_loop = uid_loop_num[INSN_UID (XEXP (x, 0))];
2556
2557 /* Link together all labels that branch outside the loop. This
2558 is used by final_[bg]iv_value and the loop unrolling code. Also
2559 mark this LABEL_REF so we know that this branch should predict
2560 false. */
2561
2562 /* A check to make sure the label is not in an inner nested loop,
2563 since this does not count as a loop exit. */
2564 if (dest_loop != -1)
2565 {
2566 for (outer_loop = dest_loop; outer_loop != -1;
2567 outer_loop = loop_outer_loop[outer_loop])
2568 if (outer_loop == loop_num)
2569 break;
2570 }
2571 else
2572 outer_loop = -1;
2573
2574 if (loop_num != -1 && outer_loop == -1)
2575 {
2576 LABEL_OUTSIDE_LOOP_P (x) = 1;
2577 LABEL_NEXTREF (x) = loop_number_exit_labels[loop_num];
2578 loop_number_exit_labels[loop_num] = x;
2579
2580 for (outer_loop = loop_num;
2581 outer_loop != -1 && outer_loop != dest_loop;
2582 outer_loop = loop_outer_loop[outer_loop])
2583 loop_number_exit_count[outer_loop]++;
2584 }
2585
2586 /* If this is inside a loop, but not in the current loop or one enclosed
2587 by it, it invalidates at least one loop. */
2588
2589 if (dest_loop == -1)
2590 return;
2591
2592 /* We must invalidate every nested loop containing the target of this
2593 label, except those that also contain the jump insn. */
2594
2595 for (; dest_loop != -1; dest_loop = loop_outer_loop[dest_loop])
2596 {
2597 /* Stop when we reach a loop that also contains the jump insn. */
2598 for (outer_loop = loop_num; outer_loop != -1;
2599 outer_loop = loop_outer_loop[outer_loop])
2600 if (dest_loop == outer_loop)
2601 return;
2602
2603 /* If we get here, we know we need to invalidate a loop. */
2604 if (loop_dump_stream && ! loop_invalid[dest_loop])
2605 fprintf (loop_dump_stream,
2606 "\nLoop at %d ignored due to multiple entry points.\n",
2607 INSN_UID (loop_number_loop_starts[dest_loop]));
2608
2609 loop_invalid[dest_loop] = 1;
2610 }
2611 return;
2612
2613 case SET:
2614 /* If this is not setting pc, ignore. */
2615 if (SET_DEST (x) == pc_rtx)
2616 mark_loop_jump (SET_SRC (x), loop_num);
2617 return;
2618
2619 case IF_THEN_ELSE:
2620 mark_loop_jump (XEXP (x, 1), loop_num);
2621 mark_loop_jump (XEXP (x, 2), loop_num);
2622 return;
2623
2624 case PARALLEL:
2625 case ADDR_VEC:
2626 for (i = 0; i < XVECLEN (x, 0); i++)
2627 mark_loop_jump (XVECEXP (x, 0, i), loop_num);
2628 return;
2629
2630 case ADDR_DIFF_VEC:
2631 for (i = 0; i < XVECLEN (x, 1); i++)
2632 mark_loop_jump (XVECEXP (x, 1, i), loop_num);
2633 return;
2634
2635 default:
2636 /* Treat anything else (such as a symbol_ref)
2637 as a branch out of this loop, but not into any loop. */
2638
2639 if (loop_num != -1)
2640 {
2641 loop_number_exit_labels[loop_num] = x;
2642
2643 for (outer_loop = loop_num; outer_loop != -1;
2644 outer_loop = loop_outer_loop[outer_loop])
2645 loop_number_exit_count[outer_loop]++;
2646 }
2647 return;
2648 }
2649 }
2650 \f
2651 /* Return nonzero if there is a label in the range from
2652 insn INSN to and including the insn whose luid is END
2653 INSN must have an assigned luid (i.e., it must not have
2654 been previously created by loop.c). */
2655
2656 static int
2657 labels_in_range_p (insn, end)
2658 rtx insn;
2659 int end;
2660 {
2661 while (insn && INSN_LUID (insn) <= end)
2662 {
2663 if (GET_CODE (insn) == CODE_LABEL)
2664 return 1;
2665 insn = NEXT_INSN (insn);
2666 }
2667
2668 return 0;
2669 }
2670
2671 /* Record that a memory reference X is being set. */
2672
2673 static void
2674 note_addr_stored (x)
2675 rtx x;
2676 {
2677 register int i;
2678
2679 if (x == 0 || GET_CODE (x) != MEM)
2680 return;
2681
2682 /* Count number of memory writes.
2683 This affects heuristics in strength_reduce. */
2684 num_mem_sets++;
2685
2686 /* BLKmode MEM means all memory is clobbered. */
2687 if (GET_MODE (x) == BLKmode)
2688 unknown_address_altered = 1;
2689
2690 if (unknown_address_altered)
2691 return;
2692
2693 for (i = 0; i < loop_store_mems_idx; i++)
2694 if (rtx_equal_p (XEXP (loop_store_mems[i], 0), XEXP (x, 0))
2695 && MEM_IN_STRUCT_P (x) == MEM_IN_STRUCT_P (loop_store_mems[i]))
2696 {
2697 /* We are storing at the same address as previously noted. Save the
2698 wider reference. */
2699 if (GET_MODE_SIZE (GET_MODE (x))
2700 > GET_MODE_SIZE (GET_MODE (loop_store_mems[i])))
2701 loop_store_mems[i] = x;
2702 break;
2703 }
2704
2705 if (i == NUM_STORES)
2706 unknown_address_altered = 1;
2707
2708 else if (i == loop_store_mems_idx)
2709 loop_store_mems[loop_store_mems_idx++] = x;
2710 }
2711 \f
2712 /* Return nonzero if the rtx X is invariant over the current loop.
2713
2714 The value is 2 if we refer to something only conditionally invariant.
2715
2716 If `unknown_address_altered' is nonzero, no memory ref is invariant.
2717 Otherwise, a memory ref is invariant if it does not conflict with
2718 anything stored in `loop_store_mems'. */
2719
2720 int
2721 invariant_p (x)
2722 register rtx x;
2723 {
2724 register int i;
2725 register enum rtx_code code;
2726 register char *fmt;
2727 int conditional = 0;
2728
2729 if (x == 0)
2730 return 1;
2731 code = GET_CODE (x);
2732 switch (code)
2733 {
2734 case CONST_INT:
2735 case CONST_DOUBLE:
2736 case SYMBOL_REF:
2737 case CONST:
2738 return 1;
2739
2740 case LABEL_REF:
2741 /* A LABEL_REF is normally invariant, however, if we are unrolling
2742 loops, and this label is inside the loop, then it isn't invariant.
2743 This is because each unrolled copy of the loop body will have
2744 a copy of this label. If this was invariant, then an insn loading
2745 the address of this label into a register might get moved outside
2746 the loop, and then each loop body would end up using the same label.
2747
2748 We don't know the loop bounds here though, so just fail for all
2749 labels. */
2750 if (flag_unroll_loops)
2751 return 0;
2752 else
2753 return 1;
2754
2755 case PC:
2756 case CC0:
2757 case UNSPEC_VOLATILE:
2758 return 0;
2759
2760 case REG:
2761 /* We used to check RTX_UNCHANGING_P (x) here, but that is invalid
2762 since the reg might be set by initialization within the loop. */
2763 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx
2764 || x == arg_pointer_rtx)
2765 return 1;
2766 if (loop_has_call
2767 && REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)])
2768 return 0;
2769 if (n_times_set[REGNO (x)] < 0)
2770 return 2;
2771 return n_times_set[REGNO (x)] == 0;
2772
2773 case MEM:
2774 /* Volatile memory references must be rejected. Do this before
2775 checking for read-only items, so that volatile read-only items
2776 will be rejected also. */
2777 if (MEM_VOLATILE_P (x))
2778 return 0;
2779
2780 /* Read-only items (such as constants in a constant pool) are
2781 invariant if their address is. */
2782 if (RTX_UNCHANGING_P (x))
2783 break;
2784
2785 /* If we filled the table (or had a subroutine call), any location
2786 in memory could have been clobbered. */
2787 if (unknown_address_altered)
2788 return 0;
2789
2790 /* See if there is any dependence between a store and this load. */
2791 for (i = loop_store_mems_idx - 1; i >= 0; i--)
2792 if (true_dependence (loop_store_mems[i], x))
2793 return 0;
2794
2795 /* It's not invalidated by a store in memory
2796 but we must still verify the address is invariant. */
2797 break;
2798
2799 case ASM_OPERANDS:
2800 /* Don't mess with insns declared volatile. */
2801 if (MEM_VOLATILE_P (x))
2802 return 0;
2803 }
2804
2805 fmt = GET_RTX_FORMAT (code);
2806 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2807 {
2808 if (fmt[i] == 'e')
2809 {
2810 int tem = invariant_p (XEXP (x, i));
2811 if (tem == 0)
2812 return 0;
2813 if (tem == 2)
2814 conditional = 1;
2815 }
2816 else if (fmt[i] == 'E')
2817 {
2818 register int j;
2819 for (j = 0; j < XVECLEN (x, i); j++)
2820 {
2821 int tem = invariant_p (XVECEXP (x, i, j));
2822 if (tem == 0)
2823 return 0;
2824 if (tem == 2)
2825 conditional = 1;
2826 }
2827
2828 }
2829 }
2830
2831 return 1 + conditional;
2832 }
2833
2834 \f
2835 /* Return nonzero if all the insns in the loop that set REG
2836 are INSN and the immediately following insns,
2837 and if each of those insns sets REG in an invariant way
2838 (not counting uses of REG in them).
2839
2840 The value is 2 if some of these insns are only conditionally invariant.
2841
2842 We assume that INSN itself is the first set of REG
2843 and that its source is invariant. */
2844
2845 static int
2846 consec_sets_invariant_p (reg, n_sets, insn)
2847 int n_sets;
2848 rtx reg, insn;
2849 {
2850 register rtx p = insn;
2851 register int regno = REGNO (reg);
2852 rtx temp;
2853 /* Number of sets we have to insist on finding after INSN. */
2854 int count = n_sets - 1;
2855 int old = n_times_set[regno];
2856 int value = 0;
2857 int this;
2858
2859 /* If N_SETS hit the limit, we can't rely on its value. */
2860 if (n_sets == 127)
2861 return 0;
2862
2863 n_times_set[regno] = 0;
2864
2865 while (count > 0)
2866 {
2867 register enum rtx_code code;
2868 rtx set;
2869
2870 p = NEXT_INSN (p);
2871 code = GET_CODE (p);
2872
2873 /* If library call, skip to end of of it. */
2874 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
2875 p = XEXP (temp, 0);
2876
2877 this = 0;
2878 if (code == INSN
2879 && (set = single_set (p))
2880 && GET_CODE (SET_DEST (set)) == REG
2881 && REGNO (SET_DEST (set)) == regno)
2882 {
2883 this = invariant_p (SET_SRC (set));
2884 if (this != 0)
2885 value |= this;
2886 else if (temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
2887 {
2888 /* If this is a libcall, then any invariant REG_EQUAL note is OK.
2889 If this is an ordinary insn, then only CONSTANT_P REG_EQUAL
2890 notes are OK. */
2891 this = (CONSTANT_P (XEXP (temp, 0))
2892 || (find_reg_note (p, REG_RETVAL, NULL_RTX)
2893 && invariant_p (XEXP (temp, 0))));
2894 if (this != 0)
2895 value |= this;
2896 }
2897 }
2898 if (this != 0)
2899 count--;
2900 else if (code != NOTE)
2901 {
2902 n_times_set[regno] = old;
2903 return 0;
2904 }
2905 }
2906
2907 n_times_set[regno] = old;
2908 /* If invariant_p ever returned 2, we return 2. */
2909 return 1 + (value & 2);
2910 }
2911
2912 #if 0
2913 /* I don't think this condition is sufficient to allow INSN
2914 to be moved, so we no longer test it. */
2915
2916 /* Return 1 if all insns in the basic block of INSN and following INSN
2917 that set REG are invariant according to TABLE. */
2918
2919 static int
2920 all_sets_invariant_p (reg, insn, table)
2921 rtx reg, insn;
2922 short *table;
2923 {
2924 register rtx p = insn;
2925 register int regno = REGNO (reg);
2926
2927 while (1)
2928 {
2929 register enum rtx_code code;
2930 p = NEXT_INSN (p);
2931 code = GET_CODE (p);
2932 if (code == CODE_LABEL || code == JUMP_INSN)
2933 return 1;
2934 if (code == INSN && GET_CODE (PATTERN (p)) == SET
2935 && GET_CODE (SET_DEST (PATTERN (p))) == REG
2936 && REGNO (SET_DEST (PATTERN (p))) == regno)
2937 {
2938 if (!invariant_p (SET_SRC (PATTERN (p)), table))
2939 return 0;
2940 }
2941 }
2942 }
2943 #endif /* 0 */
2944 \f
2945 /* Look at all uses (not sets) of registers in X. For each, if it is
2946 the single use, set USAGE[REGNO] to INSN; if there was a previous use in
2947 a different insn, set USAGE[REGNO] to const0_rtx. */
2948
2949 static void
2950 find_single_use_in_loop (insn, x, usage)
2951 rtx insn;
2952 rtx x;
2953 rtx *usage;
2954 {
2955 enum rtx_code code = GET_CODE (x);
2956 char *fmt = GET_RTX_FORMAT (code);
2957 int i, j;
2958
2959 if (code == REG)
2960 usage[REGNO (x)]
2961 = (usage[REGNO (x)] != 0 && usage[REGNO (x)] != insn)
2962 ? const0_rtx : insn;
2963
2964 else if (code == SET)
2965 {
2966 /* Don't count SET_DEST if it is a REG; otherwise count things
2967 in SET_DEST because if a register is partially modified, it won't
2968 show up as a potential movable so we don't care how USAGE is set
2969 for it. */
2970 if (GET_CODE (SET_DEST (x)) != REG)
2971 find_single_use_in_loop (insn, SET_DEST (x), usage);
2972 find_single_use_in_loop (insn, SET_SRC (x), usage);
2973 }
2974 else
2975 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2976 {
2977 if (fmt[i] == 'e' && XEXP (x, i) != 0)
2978 find_single_use_in_loop (insn, XEXP (x, i), usage);
2979 else if (fmt[i] == 'E')
2980 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
2981 find_single_use_in_loop (insn, XVECEXP (x, i, j), usage);
2982 }
2983 }
2984 \f
2985 /* Increment N_TIMES_SET at the index of each register
2986 that is modified by an insn between FROM and TO.
2987 If the value of an element of N_TIMES_SET becomes 127 or more,
2988 stop incrementing it, to avoid overflow.
2989
2990 Store in SINGLE_USAGE[I] the single insn in which register I is
2991 used, if it is only used once. Otherwise, it is set to 0 (for no
2992 uses) or const0_rtx for more than one use. This parameter may be zero,
2993 in which case this processing is not done.
2994
2995 Store in *COUNT_PTR the number of actual instruction
2996 in the loop. We use this to decide what is worth moving out. */
2997
2998 /* last_set[n] is nonzero iff reg n has been set in the current basic block.
2999 In that case, it is the insn that last set reg n. */
3000
3001 static void
3002 count_loop_regs_set (from, to, may_not_move, single_usage, count_ptr, nregs)
3003 register rtx from, to;
3004 char *may_not_move;
3005 rtx *single_usage;
3006 int *count_ptr;
3007 int nregs;
3008 {
3009 register rtx *last_set = (rtx *) alloca (nregs * sizeof (rtx));
3010 register rtx insn;
3011 register int count = 0;
3012 register rtx dest;
3013
3014 bzero ((char *) last_set, nregs * sizeof (rtx));
3015 for (insn = from; insn != to; insn = NEXT_INSN (insn))
3016 {
3017 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
3018 {
3019 ++count;
3020
3021 /* If requested, record registers that have exactly one use. */
3022 if (single_usage)
3023 {
3024 find_single_use_in_loop (insn, PATTERN (insn), single_usage);
3025
3026 /* Include uses in REG_EQUAL notes. */
3027 if (REG_NOTES (insn))
3028 find_single_use_in_loop (insn, REG_NOTES (insn), single_usage);
3029 }
3030
3031 if (GET_CODE (PATTERN (insn)) == CLOBBER
3032 && GET_CODE (XEXP (PATTERN (insn), 0)) == REG)
3033 /* Don't move a reg that has an explicit clobber.
3034 We might do so sometimes, but it's not worth the pain. */
3035 may_not_move[REGNO (XEXP (PATTERN (insn), 0))] = 1;
3036
3037 if (GET_CODE (PATTERN (insn)) == SET
3038 || GET_CODE (PATTERN (insn)) == CLOBBER)
3039 {
3040 dest = SET_DEST (PATTERN (insn));
3041 while (GET_CODE (dest) == SUBREG
3042 || GET_CODE (dest) == ZERO_EXTRACT
3043 || GET_CODE (dest) == SIGN_EXTRACT
3044 || GET_CODE (dest) == STRICT_LOW_PART)
3045 dest = XEXP (dest, 0);
3046 if (GET_CODE (dest) == REG)
3047 {
3048 register int regno = REGNO (dest);
3049 /* If this is the first setting of this reg
3050 in current basic block, and it was set before,
3051 it must be set in two basic blocks, so it cannot
3052 be moved out of the loop. */
3053 if (n_times_set[regno] > 0 && last_set[regno] == 0)
3054 may_not_move[regno] = 1;
3055 /* If this is not first setting in current basic block,
3056 see if reg was used in between previous one and this.
3057 If so, neither one can be moved. */
3058 if (last_set[regno] != 0
3059 && reg_used_between_p (dest, last_set[regno], insn))
3060 may_not_move[regno] = 1;
3061 if (n_times_set[regno] < 127)
3062 ++n_times_set[regno];
3063 last_set[regno] = insn;
3064 }
3065 }
3066 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
3067 {
3068 register int i;
3069 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
3070 {
3071 register rtx x = XVECEXP (PATTERN (insn), 0, i);
3072 if (GET_CODE (x) == CLOBBER && GET_CODE (XEXP (x, 0)) == REG)
3073 /* Don't move a reg that has an explicit clobber.
3074 It's not worth the pain to try to do it correctly. */
3075 may_not_move[REGNO (XEXP (x, 0))] = 1;
3076
3077 if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
3078 {
3079 dest = SET_DEST (x);
3080 while (GET_CODE (dest) == SUBREG
3081 || GET_CODE (dest) == ZERO_EXTRACT
3082 || GET_CODE (dest) == SIGN_EXTRACT
3083 || GET_CODE (dest) == STRICT_LOW_PART)
3084 dest = XEXP (dest, 0);
3085 if (GET_CODE (dest) == REG)
3086 {
3087 register int regno = REGNO (dest);
3088 if (n_times_set[regno] > 0 && last_set[regno] == 0)
3089 may_not_move[regno] = 1;
3090 if (last_set[regno] != 0
3091 && reg_used_between_p (dest, last_set[regno], insn))
3092 may_not_move[regno] = 1;
3093 if (n_times_set[regno] < 127)
3094 ++n_times_set[regno];
3095 last_set[regno] = insn;
3096 }
3097 }
3098 }
3099 }
3100 }
3101
3102 if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN)
3103 bzero ((char *) last_set, nregs * sizeof (rtx));
3104 }
3105 *count_ptr = count;
3106 }
3107 \f
3108 /* Given a loop that is bounded by LOOP_START and LOOP_END
3109 and that is entered at SCAN_START,
3110 return 1 if the register set in SET contained in insn INSN is used by
3111 any insn that precedes INSN in cyclic order starting
3112 from the loop entry point.
3113
3114 We don't want to use INSN_LUID here because if we restrict INSN to those
3115 that have a valid INSN_LUID, it means we cannot move an invariant out
3116 from an inner loop past two loops. */
3117
3118 static int
3119 loop_reg_used_before_p (set, insn, loop_start, scan_start, loop_end)
3120 rtx set, insn, loop_start, scan_start, loop_end;
3121 {
3122 rtx reg = SET_DEST (set);
3123 rtx p;
3124
3125 /* Scan forward checking for register usage. If we hit INSN, we
3126 are done. Otherwise, if we hit LOOP_END, wrap around to LOOP_START. */
3127 for (p = scan_start; p != insn; p = NEXT_INSN (p))
3128 {
3129 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
3130 && reg_overlap_mentioned_p (reg, PATTERN (p)))
3131 return 1;
3132
3133 if (p == loop_end)
3134 p = loop_start;
3135 }
3136
3137 return 0;
3138 }
3139 \f
3140 /* A "basic induction variable" or biv is a pseudo reg that is set
3141 (within this loop) only by incrementing or decrementing it. */
3142 /* A "general induction variable" or giv is a pseudo reg whose
3143 value is a linear function of a biv. */
3144
3145 /* Bivs are recognized by `basic_induction_var';
3146 Givs by `general_induct_var'. */
3147
3148 /* Indexed by register number, indicates whether or not register is an
3149 induction variable, and if so what type. */
3150
3151 enum iv_mode *reg_iv_type;
3152
3153 /* Indexed by register number, contains pointer to `struct induction'
3154 if register is an induction variable. This holds general info for
3155 all induction variables. */
3156
3157 struct induction **reg_iv_info;
3158
3159 /* Indexed by register number, contains pointer to `struct iv_class'
3160 if register is a basic induction variable. This holds info describing
3161 the class (a related group) of induction variables that the biv belongs
3162 to. */
3163
3164 struct iv_class **reg_biv_class;
3165
3166 /* The head of a list which links together (via the next field)
3167 every iv class for the current loop. */
3168
3169 struct iv_class *loop_iv_list;
3170
3171 /* Communication with routines called via `note_stores'. */
3172
3173 static rtx note_insn;
3174
3175 /* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */
3176
3177 static rtx addr_placeholder;
3178
3179 /* ??? Unfinished optimizations, and possible future optimizations,
3180 for the strength reduction code. */
3181
3182 /* ??? There is one more optimization you might be interested in doing: to
3183 allocate pseudo registers for frequently-accessed memory locations.
3184 If the same memory location is referenced each time around, it might
3185 be possible to copy it into a register before and out after.
3186 This is especially useful when the memory location is a variable which
3187 is in a stack slot because somewhere its address is taken. If the
3188 loop doesn't contain a function call and the variable isn't volatile,
3189 it is safe to keep the value in a register for the duration of the
3190 loop. One tricky thing is that the copying of the value back from the
3191 register has to be done on all exits from the loop. You need to check that
3192 all the exits from the loop go to the same place. */
3193
3194 /* ??? The interaction of biv elimination, and recognition of 'constant'
3195 bivs, may cause problems. */
3196
3197 /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
3198 performance problems.
3199
3200 Perhaps don't eliminate things that can be combined with an addressing
3201 mode. Find all givs that have the same biv, mult_val, and add_val;
3202 then for each giv, check to see if its only use dies in a following
3203 memory address. If so, generate a new memory address and check to see
3204 if it is valid. If it is valid, then store the modified memory address,
3205 otherwise, mark the giv as not done so that it will get its own iv. */
3206
3207 /* ??? Could try to optimize branches when it is known that a biv is always
3208 positive. */
3209
3210 /* ??? When replace a biv in a compare insn, we should replace with closest
3211 giv so that an optimized branch can still be recognized by the combiner,
3212 e.g. the VAX acb insn. */
3213
3214 /* ??? Many of the checks involving uid_luid could be simplified if regscan
3215 was rerun in loop_optimize whenever a register was added or moved.
3216 Also, some of the optimizations could be a little less conservative. */
3217 \f
3218 /* Perform strength reduction and induction variable elimination. */
3219
3220 /* Pseudo registers created during this function will be beyond the last
3221 valid index in several tables including n_times_set and regno_last_uid.
3222 This does not cause a problem here, because the added registers cannot be
3223 givs outside of their loop, and hence will never be reconsidered.
3224 But scan_loop must check regnos to make sure they are in bounds. */
3225
3226 static void
3227 strength_reduce (scan_start, end, loop_top, insn_count,
3228 loop_start, loop_end)
3229 rtx scan_start;
3230 rtx end;
3231 rtx loop_top;
3232 int insn_count;
3233 rtx loop_start;
3234 rtx loop_end;
3235 {
3236 rtx p;
3237 rtx set;
3238 rtx inc_val;
3239 rtx mult_val;
3240 rtx dest_reg;
3241 /* This is 1 if current insn is not executed at least once for every loop
3242 iteration. */
3243 int not_every_iteration = 0;
3244 /* This is 1 if current insn may be executed more than once for every
3245 loop iteration. */
3246 int maybe_multiple = 0;
3247 /* Temporary list pointers for traversing loop_iv_list. */
3248 struct iv_class *bl, **backbl;
3249 /* Ratio of extra register life span we can justify
3250 for saving an instruction. More if loop doesn't call subroutines
3251 since in that case saving an insn makes more difference
3252 and more registers are available. */
3253 /* ??? could set this to last value of threshold in move_movables */
3254 int threshold = (loop_has_call ? 1 : 2) * (3 + n_non_fixed_regs);
3255 /* Map of pseudo-register replacements. */
3256 rtx *reg_map;
3257 int call_seen;
3258 rtx test;
3259 rtx end_insert_before;
3260 int loop_depth = 0;
3261
3262 reg_iv_type = (enum iv_mode *) alloca (max_reg_before_loop
3263 * sizeof (enum iv_mode *));
3264 bzero ((char *) reg_iv_type, max_reg_before_loop * sizeof (enum iv_mode *));
3265 reg_iv_info = (struct induction **)
3266 alloca (max_reg_before_loop * sizeof (struct induction *));
3267 bzero ((char *) reg_iv_info, (max_reg_before_loop
3268 * sizeof (struct induction *)));
3269 reg_biv_class = (struct iv_class **)
3270 alloca (max_reg_before_loop * sizeof (struct iv_class *));
3271 bzero ((char *) reg_biv_class, (max_reg_before_loop
3272 * sizeof (struct iv_class *)));
3273
3274 loop_iv_list = 0;
3275 addr_placeholder = gen_reg_rtx (Pmode);
3276
3277 /* Save insn immediately after the loop_end. Insns inserted after loop_end
3278 must be put before this insn, so that they will appear in the right
3279 order (i.e. loop order).
3280
3281 If loop_end is the end of the current function, then emit a
3282 NOTE_INSN_DELETED after loop_end and set end_insert_before to the
3283 dummy note insn. */
3284 if (NEXT_INSN (loop_end) != 0)
3285 end_insert_before = NEXT_INSN (loop_end);
3286 else
3287 end_insert_before = emit_note_after (NOTE_INSN_DELETED, loop_end);
3288
3289 /* Scan through loop to find all possible bivs. */
3290
3291 p = scan_start;
3292 while (1)
3293 {
3294 p = NEXT_INSN (p);
3295 /* At end of a straight-in loop, we are done.
3296 At end of a loop entered at the bottom, scan the top. */
3297 if (p == scan_start)
3298 break;
3299 if (p == end)
3300 {
3301 if (loop_top != 0)
3302 p = loop_top;
3303 else
3304 break;
3305 if (p == scan_start)
3306 break;
3307 }
3308
3309 if (GET_CODE (p) == INSN
3310 && (set = single_set (p))
3311 && GET_CODE (SET_DEST (set)) == REG)
3312 {
3313 dest_reg = SET_DEST (set);
3314 if (REGNO (dest_reg) < max_reg_before_loop
3315 && REGNO (dest_reg) >= FIRST_PSEUDO_REGISTER
3316 && reg_iv_type[REGNO (dest_reg)] != NOT_BASIC_INDUCT)
3317 {
3318 if (basic_induction_var (SET_SRC (set), GET_MODE (SET_SRC (set)),
3319 dest_reg, p, &inc_val, &mult_val))
3320 {
3321 /* It is a possible basic induction variable.
3322 Create and initialize an induction structure for it. */
3323
3324 struct induction *v
3325 = (struct induction *) alloca (sizeof (struct induction));
3326
3327 record_biv (v, p, dest_reg, inc_val, mult_val,
3328 not_every_iteration, maybe_multiple);
3329 reg_iv_type[REGNO (dest_reg)] = BASIC_INDUCT;
3330 }
3331 else if (REGNO (dest_reg) < max_reg_before_loop)
3332 reg_iv_type[REGNO (dest_reg)] = NOT_BASIC_INDUCT;
3333 }
3334 }
3335
3336 /* Past CODE_LABEL, we get to insns that may be executed multiple
3337 times. The only way we can be sure that they can't is if every
3338 every jump insn between here and the end of the loop either
3339 returns, exits the loop, is a forward jump, or is a jump
3340 to the loop start. */
3341
3342 if (GET_CODE (p) == CODE_LABEL)
3343 {
3344 rtx insn = p;
3345
3346 maybe_multiple = 0;
3347
3348 while (1)
3349 {
3350 insn = NEXT_INSN (insn);
3351 if (insn == scan_start)
3352 break;
3353 if (insn == end)
3354 {
3355 if (loop_top != 0)
3356 insn = loop_top;
3357 else
3358 break;
3359 if (insn == scan_start)
3360 break;
3361 }
3362
3363 if (GET_CODE (insn) == JUMP_INSN
3364 && GET_CODE (PATTERN (insn)) != RETURN
3365 && (! condjump_p (insn)
3366 || (JUMP_LABEL (insn) != 0
3367 && JUMP_LABEL (insn) != scan_start
3368 && (INSN_UID (JUMP_LABEL (insn)) >= max_uid_for_loop
3369 || INSN_UID (insn) >= max_uid_for_loop
3370 || (INSN_LUID (JUMP_LABEL (insn))
3371 < INSN_LUID (insn))))))
3372 {
3373 maybe_multiple = 1;
3374 break;
3375 }
3376 }
3377 }
3378
3379 /* Past a jump, we get to insns for which we can't count
3380 on whether they will be executed during each iteration. */
3381 /* This code appears twice in strength_reduce. There is also similar
3382 code in scan_loop. */
3383 if (GET_CODE (p) == JUMP_INSN
3384 /* If we enter the loop in the middle, and scan around to the
3385 beginning, don't set not_every_iteration for that.
3386 This can be any kind of jump, since we want to know if insns
3387 will be executed if the loop is executed. */
3388 && ! (JUMP_LABEL (p) == loop_top
3389 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3390 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3391 {
3392 rtx label = 0;
3393
3394 /* If this is a jump outside the loop, then it also doesn't
3395 matter. Check to see if the target of this branch is on the
3396 loop_number_exits_labels list. */
3397
3398 for (label = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]];
3399 label;
3400 label = LABEL_NEXTREF (label))
3401 if (XEXP (label, 0) == JUMP_LABEL (p))
3402 break;
3403
3404 if (! label)
3405 not_every_iteration = 1;
3406 }
3407
3408 else if (GET_CODE (p) == NOTE)
3409 {
3410 /* At the virtual top of a converted loop, insns are again known to
3411 be executed each iteration: logically, the loop begins here
3412 even though the exit code has been duplicated. */
3413 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
3414 not_every_iteration = 0;
3415 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
3416 loop_depth++;
3417 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
3418 loop_depth--;
3419 }
3420
3421 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3422 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3423 or not an insn is known to be executed each iteration of the
3424 loop, whether or not any iterations are known to occur.
3425
3426 Therefore, if we have just passed a label and have no more labels
3427 between here and the test insn of the loop, we know these insns
3428 will be executed each iteration. */
3429
3430 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
3431 && no_labels_between_p (p, loop_end))
3432 not_every_iteration = 0;
3433 }
3434
3435 /* Scan loop_iv_list to remove all regs that proved not to be bivs.
3436 Make a sanity check against n_times_set. */
3437 for (backbl = &loop_iv_list, bl = *backbl; bl; bl = bl->next)
3438 {
3439 if (reg_iv_type[bl->regno] != BASIC_INDUCT
3440 /* Above happens if register modified by subreg, etc. */
3441 /* Make sure it is not recognized as a basic induction var: */
3442 || n_times_set[bl->regno] != bl->biv_count
3443 /* If never incremented, it is invariant that we decided not to
3444 move. So leave it alone. */
3445 || ! bl->incremented)
3446 {
3447 if (loop_dump_stream)
3448 fprintf (loop_dump_stream, "Reg %d: biv discarded, %s\n",
3449 bl->regno,
3450 (reg_iv_type[bl->regno] != BASIC_INDUCT
3451 ? "not induction variable"
3452 : (! bl->incremented ? "never incremented"
3453 : "count error")));
3454
3455 reg_iv_type[bl->regno] = NOT_BASIC_INDUCT;
3456 *backbl = bl->next;
3457 }
3458 else
3459 {
3460 backbl = &bl->next;
3461
3462 if (loop_dump_stream)
3463 fprintf (loop_dump_stream, "Reg %d: biv verified\n", bl->regno);
3464 }
3465 }
3466
3467 /* Exit if there are no bivs. */
3468 if (! loop_iv_list)
3469 {
3470 /* Can still unroll the loop anyways, but indicate that there is no
3471 strength reduction info available. */
3472 if (flag_unroll_loops)
3473 unroll_loop (loop_end, insn_count, loop_start, end_insert_before, 0);
3474
3475 return;
3476 }
3477
3478 /* Find initial value for each biv by searching backwards from loop_start,
3479 halting at first label. Also record any test condition. */
3480
3481 call_seen = 0;
3482 for (p = loop_start; p && GET_CODE (p) != CODE_LABEL; p = PREV_INSN (p))
3483 {
3484 note_insn = p;
3485
3486 if (GET_CODE (p) == CALL_INSN)
3487 call_seen = 1;
3488
3489 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3490 || GET_CODE (p) == CALL_INSN)
3491 note_stores (PATTERN (p), record_initial);
3492
3493 /* Record any test of a biv that branches around the loop if no store
3494 between it and the start of loop. We only care about tests with
3495 constants and registers and only certain of those. */
3496 if (GET_CODE (p) == JUMP_INSN
3497 && JUMP_LABEL (p) != 0
3498 && next_real_insn (JUMP_LABEL (p)) == next_real_insn (loop_end)
3499 && (test = get_condition_for_loop (p)) != 0
3500 && GET_CODE (XEXP (test, 0)) == REG
3501 && REGNO (XEXP (test, 0)) < max_reg_before_loop
3502 && (bl = reg_biv_class[REGNO (XEXP (test, 0))]) != 0
3503 && valid_initial_value_p (XEXP (test, 1), p, call_seen, loop_start)
3504 && bl->init_insn == 0)
3505 {
3506 /* If an NE test, we have an initial value! */
3507 if (GET_CODE (test) == NE)
3508 {
3509 bl->init_insn = p;
3510 bl->init_set = gen_rtx (SET, VOIDmode,
3511 XEXP (test, 0), XEXP (test, 1));
3512 }
3513 else
3514 bl->initial_test = test;
3515 }
3516 }
3517
3518 /* Look at the each biv and see if we can say anything better about its
3519 initial value from any initializing insns set up above. (This is done
3520 in two passes to avoid missing SETs in a PARALLEL.) */
3521 for (bl = loop_iv_list; bl; bl = bl->next)
3522 {
3523 rtx src;
3524
3525 if (! bl->init_insn)
3526 continue;
3527
3528 src = SET_SRC (bl->init_set);
3529
3530 if (loop_dump_stream)
3531 fprintf (loop_dump_stream,
3532 "Biv %d initialized at insn %d: initial value ",
3533 bl->regno, INSN_UID (bl->init_insn));
3534
3535 if ((GET_MODE (src) == GET_MODE (regno_reg_rtx[bl->regno])
3536 || GET_MODE (src) == VOIDmode)
3537 && valid_initial_value_p (src, bl->init_insn, call_seen, loop_start))
3538 {
3539 bl->initial_value = src;
3540
3541 if (loop_dump_stream)
3542 {
3543 if (GET_CODE (src) == CONST_INT)
3544 fprintf (loop_dump_stream, "%d\n", INTVAL (src));
3545 else
3546 {
3547 print_rtl (loop_dump_stream, src);
3548 fprintf (loop_dump_stream, "\n");
3549 }
3550 }
3551 }
3552 else
3553 {
3554 /* Biv initial value is not simple move,
3555 so let it keep initial value of "itself". */
3556
3557 if (loop_dump_stream)
3558 fprintf (loop_dump_stream, "is complex\n");
3559 }
3560 }
3561
3562 /* Search the loop for general induction variables. */
3563
3564 /* A register is a giv if: it is only set once, it is a function of a
3565 biv and a constant (or invariant), and it is not a biv. */
3566
3567 not_every_iteration = 0;
3568 loop_depth = 0;
3569 p = scan_start;
3570 while (1)
3571 {
3572 p = NEXT_INSN (p);
3573 /* At end of a straight-in loop, we are done.
3574 At end of a loop entered at the bottom, scan the top. */
3575 if (p == scan_start)
3576 break;
3577 if (p == end)
3578 {
3579 if (loop_top != 0)
3580 p = loop_top;
3581 else
3582 break;
3583 if (p == scan_start)
3584 break;
3585 }
3586
3587 /* Look for a general induction variable in a register. */
3588 if (GET_CODE (p) == INSN
3589 && (set = single_set (p))
3590 && GET_CODE (SET_DEST (set)) == REG
3591 && ! may_not_optimize[REGNO (SET_DEST (set))])
3592 {
3593 rtx src_reg;
3594 rtx add_val;
3595 rtx mult_val;
3596 int benefit;
3597 rtx regnote = 0;
3598
3599 dest_reg = SET_DEST (set);
3600 if (REGNO (dest_reg) < FIRST_PSEUDO_REGISTER)
3601 continue;
3602
3603 if (/* SET_SRC is a giv. */
3604 ((benefit = general_induction_var (SET_SRC (set),
3605 &src_reg, &add_val,
3606 &mult_val))
3607 /* Equivalent expression is a giv. */
3608 || ((regnote = find_reg_note (p, REG_EQUAL, NULL_RTX))
3609 && (benefit = general_induction_var (XEXP (regnote, 0),
3610 &src_reg,
3611 &add_val, &mult_val))))
3612 /* Don't try to handle any regs made by loop optimization.
3613 We have nothing on them in regno_first_uid, etc. */
3614 && REGNO (dest_reg) < max_reg_before_loop
3615 /* Don't recognize a BASIC_INDUCT_VAR here. */
3616 && dest_reg != src_reg
3617 /* This must be the only place where the register is set. */
3618 && (n_times_set[REGNO (dest_reg)] == 1
3619 /* or all sets must be consecutive and make a giv. */
3620 || (benefit = consec_sets_giv (benefit, p,
3621 src_reg, dest_reg,
3622 &add_val, &mult_val))))
3623 {
3624 int count;
3625 struct induction *v
3626 = (struct induction *) alloca (sizeof (struct induction));
3627 rtx temp;
3628
3629 /* If this is a library call, increase benefit. */
3630 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
3631 benefit += libcall_benefit (p);
3632
3633 /* Skip the consecutive insns, if there are any. */
3634 for (count = n_times_set[REGNO (dest_reg)] - 1;
3635 count > 0; count--)
3636 {
3637 /* If first insn of libcall sequence, skip to end.
3638 Do this at start of loop, since INSN is guaranteed to
3639 be an insn here. */
3640 if (GET_CODE (p) != NOTE
3641 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
3642 p = XEXP (temp, 0);
3643
3644 do p = NEXT_INSN (p);
3645 while (GET_CODE (p) == NOTE);
3646 }
3647
3648 record_giv (v, p, src_reg, dest_reg, mult_val, add_val, benefit,
3649 DEST_REG, not_every_iteration, NULL_PTR, loop_start,
3650 loop_end);
3651
3652 }
3653 }
3654
3655 #ifndef DONT_REDUCE_ADDR
3656 /* Look for givs which are memory addresses. */
3657 /* This resulted in worse code on a VAX 8600. I wonder if it
3658 still does. */
3659 if (GET_CODE (p) == INSN)
3660 find_mem_givs (PATTERN (p), p, not_every_iteration, loop_start,
3661 loop_end);
3662 #endif
3663
3664 /* Update the status of whether giv can derive other givs. This can
3665 change when we pass a label or an insn that updates a biv. */
3666 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3667 || GET_CODE (p) == CODE_LABEL)
3668 update_giv_derive (p);
3669
3670 /* Past a jump, we get to insns for which we can't count
3671 on whether they will be executed during each iteration. */
3672 /* This code appears twice in strength_reduce. There is also similar
3673 code in scan_loop. */
3674 if (GET_CODE (p) == JUMP_INSN
3675 /* If we enter the loop in the middle, and scan around to the
3676 beginning, don't set not_every_iteration for that.
3677 This can be any kind of jump, since we want to know if insns
3678 will be executed if the loop is executed. */
3679 && ! (JUMP_LABEL (p) == loop_top
3680 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3681 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3682 {
3683 rtx label = 0;
3684
3685 /* If this is a jump outside the loop, then it also doesn't
3686 matter. Check to see if the target of this branch is on the
3687 loop_number_exits_labels list. */
3688
3689 for (label = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]];
3690 label;
3691 label = LABEL_NEXTREF (label))
3692 if (XEXP (label, 0) == JUMP_LABEL (p))
3693 break;
3694
3695 if (! label)
3696 not_every_iteration = 1;
3697 }
3698
3699 else if (GET_CODE (p) == NOTE)
3700 {
3701 /* At the virtual top of a converted loop, insns are again known to
3702 be executed each iteration: logically, the loop begins here
3703 even though the exit code has been duplicated. */
3704 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
3705 not_every_iteration = 0;
3706 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
3707 loop_depth++;
3708 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
3709 loop_depth--;
3710 }
3711
3712 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3713 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3714 or not an insn is known to be executed each iteration of the
3715 loop, whether or not any iterations are known to occur.
3716
3717 Therefore, if we have just passed a label and have no more labels
3718 between here and the test insn of the loop, we know these insns
3719 will be executed each iteration. */
3720
3721 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
3722 && no_labels_between_p (p, loop_end))
3723 not_every_iteration = 0;
3724 }
3725
3726 /* Try to calculate and save the number of loop iterations. This is
3727 set to zero if the actual number can not be calculated. This must
3728 be called after all giv's have been identified, since otherwise it may
3729 fail if the iteration variable is a giv. */
3730
3731 loop_n_iterations = loop_iterations (loop_start, loop_end);
3732
3733 /* Now for each giv for which we still don't know whether or not it is
3734 replaceable, check to see if it is replaceable because its final value
3735 can be calculated. This must be done after loop_iterations is called,
3736 so that final_giv_value will work correctly. */
3737
3738 for (bl = loop_iv_list; bl; bl = bl->next)
3739 {
3740 struct induction *v;
3741
3742 for (v = bl->giv; v; v = v->next_iv)
3743 if (! v->replaceable && ! v->not_replaceable)
3744 check_final_value (v, loop_start, loop_end);
3745 }
3746
3747 /* Try to prove that the loop counter variable (if any) is always
3748 nonnegative; if so, record that fact with a REG_NONNEG note
3749 so that "decrement and branch until zero" insn can be used. */
3750 check_dbra_loop (loop_end, insn_count, loop_start);
3751
3752 /* Create reg_map to hold substitutions for replaceable giv regs. */
3753 reg_map = (rtx *) alloca (max_reg_before_loop * sizeof (rtx));
3754 bzero ((char *) reg_map, max_reg_before_loop * sizeof (rtx));
3755
3756 /* Examine each iv class for feasibility of strength reduction/induction
3757 variable elimination. */
3758
3759 for (bl = loop_iv_list; bl; bl = bl->next)
3760 {
3761 struct induction *v;
3762 int benefit;
3763 int all_reduced;
3764 rtx final_value = 0;
3765
3766 /* Test whether it will be possible to eliminate this biv
3767 provided all givs are reduced. This is possible if either
3768 the reg is not used outside the loop, or we can compute
3769 what its final value will be.
3770
3771 For architectures with a decrement_and_branch_until_zero insn,
3772 don't do this if we put a REG_NONNEG note on the endtest for
3773 this biv. */
3774
3775 /* Compare against bl->init_insn rather than loop_start.
3776 We aren't concerned with any uses of the biv between
3777 init_insn and loop_start since these won't be affected
3778 by the value of the biv elsewhere in the function, so
3779 long as init_insn doesn't use the biv itself.
3780 March 14, 1989 -- self@bayes.arc.nasa.gov */
3781
3782 if ((uid_luid[regno_last_uid[bl->regno]] < INSN_LUID (loop_end)
3783 && bl->init_insn
3784 && INSN_UID (bl->init_insn) < max_uid_for_loop
3785 && uid_luid[regno_first_uid[bl->regno]] >= INSN_LUID (bl->init_insn)
3786 #ifdef HAVE_decrement_and_branch_until_zero
3787 && ! bl->nonneg
3788 #endif
3789 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
3790 || ((final_value = final_biv_value (bl, loop_start, loop_end))
3791 #ifdef HAVE_decrement_and_branch_until_zero
3792 && ! bl->nonneg
3793 #endif
3794 ))
3795 bl->eliminable = maybe_eliminate_biv (bl, loop_start, end, 0,
3796 threshold, insn_count);
3797 else
3798 {
3799 if (loop_dump_stream)
3800 {
3801 fprintf (loop_dump_stream,
3802 "Cannot eliminate biv %d.\n",
3803 bl->regno);
3804 fprintf (loop_dump_stream,
3805 "First use: insn %d, last use: insn %d.\n",
3806 regno_first_uid[bl->regno],
3807 regno_last_uid[bl->regno]);
3808 }
3809 }
3810
3811 /* Combine all giv's for this iv_class. */
3812 combine_givs (bl);
3813
3814 /* This will be true at the end, if all givs which depend on this
3815 biv have been strength reduced.
3816 We can't (currently) eliminate the biv unless this is so. */
3817 all_reduced = 1;
3818
3819 /* Check each giv in this class to see if we will benefit by reducing
3820 it. Skip giv's combined with others. */
3821 for (v = bl->giv; v; v = v->next_iv)
3822 {
3823 struct induction *tv;
3824
3825 if (v->ignore || v->same)
3826 continue;
3827
3828 benefit = v->benefit;
3829
3830 /* Reduce benefit if not replaceable, since we will insert
3831 a move-insn to replace the insn that calculates this giv.
3832 Don't do this unless the giv is a user variable, since it
3833 will often be marked non-replaceable because of the duplication
3834 of the exit code outside the loop. In such a case, the copies
3835 we insert are dead and will be deleted. So they don't have
3836 a cost. Similar situations exist. */
3837 /* ??? The new final_[bg]iv_value code does a much better job
3838 of finding replaceable giv's, and hence this code may no longer
3839 be necessary. */
3840 if (! v->replaceable && ! bl->eliminable
3841 && REG_USERVAR_P (v->dest_reg))
3842 benefit -= copy_cost;
3843
3844 /* Decrease the benefit to count the add-insns that we will
3845 insert to increment the reduced reg for the giv. */
3846 benefit -= add_cost * bl->biv_count;
3847
3848 /* Decide whether to strength-reduce this giv or to leave the code
3849 unchanged (recompute it from the biv each time it is used).
3850 This decision can be made independently for each giv. */
3851
3852 /* ??? Perhaps attempt to guess whether autoincrement will handle
3853 some of the new add insns; if so, can increase BENEFIT
3854 (undo the subtraction of add_cost that was done above). */
3855
3856 /* If an insn is not to be strength reduced, then set its ignore
3857 flag, and clear all_reduced. */
3858
3859 /* A giv that depends on a reversed biv must be reduced if it is
3860 used after the loop exit, otherwise, it would have the wrong
3861 value after the loop exit. To make it simple, just reduce all
3862 of such giv's whether or not we know they are used after the loop
3863 exit. */
3864
3865 if (v->lifetime * threshold * benefit < insn_count
3866 && ! bl->reversed)
3867 {
3868 if (loop_dump_stream)
3869 fprintf (loop_dump_stream,
3870 "giv of insn %d not worth while, %d vs %d.\n",
3871 INSN_UID (v->insn),
3872 v->lifetime * threshold * benefit, insn_count);
3873 v->ignore = 1;
3874 all_reduced = 0;
3875 }
3876 else
3877 {
3878 /* Check that we can increment the reduced giv without a
3879 multiply insn. If not, reject it. */
3880
3881 for (tv = bl->biv; tv; tv = tv->next_iv)
3882 if (tv->mult_val == const1_rtx
3883 && ! product_cheap_p (tv->add_val, v->mult_val))
3884 {
3885 if (loop_dump_stream)
3886 fprintf (loop_dump_stream,
3887 "giv of insn %d: would need a multiply.\n",
3888 INSN_UID (v->insn));
3889 v->ignore = 1;
3890 all_reduced = 0;
3891 break;
3892 }
3893 }
3894 }
3895
3896 /* Reduce each giv that we decided to reduce. */
3897
3898 for (v = bl->giv; v; v = v->next_iv)
3899 {
3900 struct induction *tv;
3901 if (! v->ignore && v->same == 0)
3902 {
3903 int auto_inc_opt = 0;
3904
3905 v->new_reg = gen_reg_rtx (v->mode);
3906
3907 #ifdef AUTO_INC_DEC
3908 /* If the target has auto-increment addressing modes, and
3909 this is an address giv, then try to put the increment
3910 immediately after its use, so that flow can create an
3911 auto-increment addressing mode. */
3912 if (v->giv_type == DEST_ADDR && bl->biv_count == 1
3913 && bl->biv->always_executed && ! bl->biv->maybe_multiple
3914 /* We don't handle reversed biv's because bl->biv->insn
3915 does not have a valid INSN_LUID. */
3916 && ! bl->reversed
3917 && v->always_executed && ! v->maybe_multiple)
3918 {
3919 /* If other giv's have been combined with this one, then
3920 this will work only if all uses of the other giv's occur
3921 before this giv's insn. This is difficult to check.
3922
3923 We simplify this by looking for the common case where
3924 there is one DEST_REG giv, and this giv's insn is the
3925 last use of the dest_reg of that DEST_REG giv. If the
3926 the increment occurs after the address giv, then we can
3927 perform the optimization. (Otherwise, the increment
3928 would have to go before other_giv, and we would not be
3929 able to combine it with the address giv to get an
3930 auto-inc address.) */
3931 if (v->combined_with)
3932 {
3933 struct induction *other_giv = 0;
3934
3935 for (tv = bl->giv; tv; tv = tv->next_iv)
3936 if (tv->same == v)
3937 {
3938 if (other_giv)
3939 break;
3940 else
3941 other_giv = tv;
3942 }
3943 if (! tv && other_giv
3944 && (regno_last_uid[REGNO (other_giv->dest_reg)]
3945 == INSN_UID (v->insn))
3946 && INSN_LUID (v->insn) < INSN_LUID (bl->biv->insn))
3947 auto_inc_opt = 1;
3948 }
3949 /* Check for case where increment is before the the address
3950 giv. */
3951 else if (INSN_LUID (v->insn) > INSN_LUID (bl->biv->insn))
3952 auto_inc_opt = -1;
3953 else
3954 auto_inc_opt = 1;
3955
3956 #ifdef HAVE_cc0
3957 {
3958 rtx prev;
3959
3960 /* We can't put an insn immediately after one setting
3961 cc0, or immediately before one using cc0. */
3962 if ((auto_inc_opt == 1 && sets_cc0_p (PATTERN (v->insn)))
3963 || (auto_inc_opt == -1
3964 && (prev = prev_nonnote_insn (v->insn)) != 0
3965 && GET_RTX_CLASS (GET_CODE (prev)) == 'i'
3966 && sets_cc0_p (PATTERN (prev))))
3967 auto_inc_opt = 0;
3968 }
3969 #endif
3970
3971 if (auto_inc_opt)
3972 v->auto_inc_opt = 1;
3973 }
3974 #endif
3975
3976 /* For each place where the biv is incremented, add an insn
3977 to increment the new, reduced reg for the giv. */
3978 for (tv = bl->biv; tv; tv = tv->next_iv)
3979 {
3980 rtx insert_before;
3981
3982 if (! auto_inc_opt)
3983 insert_before = tv->insn;
3984 else if (auto_inc_opt == 1)
3985 insert_before = NEXT_INSN (v->insn);
3986 else
3987 insert_before = v->insn;
3988
3989 if (tv->mult_val == const1_rtx)
3990 emit_iv_add_mult (tv->add_val, v->mult_val,
3991 v->new_reg, v->new_reg, insert_before);
3992 else /* tv->mult_val == const0_rtx */
3993 /* A multiply is acceptable here
3994 since this is presumed to be seldom executed. */
3995 emit_iv_add_mult (tv->add_val, v->mult_val,
3996 v->add_val, v->new_reg, insert_before);
3997 }
3998
3999 /* Add code at loop start to initialize giv's reduced reg. */
4000
4001 emit_iv_add_mult (bl->initial_value, v->mult_val,
4002 v->add_val, v->new_reg, loop_start);
4003 }
4004 }
4005
4006 /* Rescan all givs. If a giv is the same as a giv not reduced, mark it
4007 as not reduced.
4008
4009 For each giv register that can be reduced now: if replaceable,
4010 substitute reduced reg wherever the old giv occurs;
4011 else add new move insn "giv_reg = reduced_reg".
4012
4013 Also check for givs whose first use is their definition and whose
4014 last use is the definition of another giv. If so, it is likely
4015 dead and should not be used to eliminate a biv. */
4016 for (v = bl->giv; v; v = v->next_iv)
4017 {
4018 if (v->same && v->same->ignore)
4019 v->ignore = 1;
4020
4021 if (v->ignore)
4022 continue;
4023
4024 if (v->giv_type == DEST_REG
4025 && regno_first_uid[REGNO (v->dest_reg)] == INSN_UID (v->insn))
4026 {
4027 struct induction *v1;
4028
4029 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4030 if (regno_last_uid[REGNO (v->dest_reg)] == INSN_UID (v1->insn))
4031 v->maybe_dead = 1;
4032 }
4033
4034 /* Update expression if this was combined, in case other giv was
4035 replaced. */
4036 if (v->same)
4037 v->new_reg = replace_rtx (v->new_reg,
4038 v->same->dest_reg, v->same->new_reg);
4039
4040 if (v->giv_type == DEST_ADDR)
4041 /* Store reduced reg as the address in the memref where we found
4042 this giv. */
4043 validate_change (v->insn, v->location, v->new_reg, 0);
4044 else if (v->replaceable)
4045 {
4046 reg_map[REGNO (v->dest_reg)] = v->new_reg;
4047
4048 #if 0
4049 /* I can no longer duplicate the original problem. Perhaps
4050 this is unnecessary now? */
4051
4052 /* Replaceable; it isn't strictly necessary to delete the old
4053 insn and emit a new one, because v->dest_reg is now dead.
4054
4055 However, especially when unrolling loops, the special
4056 handling for (set REG0 REG1) in the second cse pass may
4057 make v->dest_reg live again. To avoid this problem, emit
4058 an insn to set the original giv reg from the reduced giv.
4059 We can not delete the original insn, since it may be part
4060 of a LIBCALL, and the code in flow that eliminates dead
4061 libcalls will fail if it is deleted. */
4062 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4063 v->insn);
4064 #endif
4065 }
4066 else
4067 {
4068 /* Not replaceable; emit an insn to set the original giv reg from
4069 the reduced giv, same as above. */
4070 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4071 v->insn);
4072 }
4073
4074 /* When a loop is reversed, givs which depend on the reversed
4075 biv, and which are live outside the loop, must be set to their
4076 correct final value. This insn is only needed if the giv is
4077 not replaceable. The correct final value is the same as the
4078 value that the giv starts the reversed loop with. */
4079 if (bl->reversed && ! v->replaceable)
4080 emit_iv_add_mult (bl->initial_value, v->mult_val,
4081 v->add_val, v->dest_reg, end_insert_before);
4082 else if (v->final_value)
4083 {
4084 rtx insert_before;
4085
4086 /* If the loop has multiple exits, emit the insn before the
4087 loop to ensure that it will always be executed no matter
4088 how the loop exits. Otherwise, emit the insn after the loop,
4089 since this is slightly more efficient. */
4090 if (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
4091 insert_before = loop_start;
4092 else
4093 insert_before = end_insert_before;
4094 emit_insn_before (gen_move_insn (v->dest_reg, v->final_value),
4095 insert_before);
4096
4097 #if 0
4098 /* If the insn to set the final value of the giv was emitted
4099 before the loop, then we must delete the insn inside the loop
4100 that sets it. If this is a LIBCALL, then we must delete
4101 every insn in the libcall. Note, however, that
4102 final_giv_value will only succeed when there are multiple
4103 exits if the giv is dead at each exit, hence it does not
4104 matter that the original insn remains because it is dead
4105 anyways. */
4106 /* Delete the insn inside the loop that sets the giv since
4107 the giv is now set before (or after) the loop. */
4108 delete_insn (v->insn);
4109 #endif
4110 }
4111
4112 if (loop_dump_stream)
4113 {
4114 fprintf (loop_dump_stream, "giv at %d reduced to ",
4115 INSN_UID (v->insn));
4116 print_rtl (loop_dump_stream, v->new_reg);
4117 fprintf (loop_dump_stream, "\n");
4118 }
4119 }
4120
4121 /* All the givs based on the biv bl have been reduced if they
4122 merit it. */
4123
4124 /* For each giv not marked as maybe dead that has been combined with a
4125 second giv, clear any "maybe dead" mark on that second giv.
4126 v->new_reg will either be or refer to the register of the giv it
4127 combined with.
4128
4129 Doing this clearing avoids problems in biv elimination where a
4130 giv's new_reg is a complex value that can't be put in the insn but
4131 the giv combined with (with a reg as new_reg) is marked maybe_dead.
4132 Since the register will be used in either case, we'd prefer it be
4133 used from the simpler giv. */
4134
4135 for (v = bl->giv; v; v = v->next_iv)
4136 if (! v->maybe_dead && v->same)
4137 v->same->maybe_dead = 0;
4138
4139 /* Try to eliminate the biv, if it is a candidate.
4140 This won't work if ! all_reduced,
4141 since the givs we planned to use might not have been reduced.
4142
4143 We have to be careful that we didn't initially think we could eliminate
4144 this biv because of a giv that we now think may be dead and shouldn't
4145 be used as a biv replacement.
4146
4147 Also, there is the possibility that we may have a giv that looks
4148 like it can be used to eliminate a biv, but the resulting insn
4149 isn't valid. This can happen, for example, on the 88k, where a
4150 JUMP_INSN can compare a register only with zero. Attempts to
4151 replace it with a compare with a constant will fail.
4152
4153 Note that in cases where this call fails, we may have replaced some
4154 of the occurrences of the biv with a giv, but no harm was done in
4155 doing so in the rare cases where it can occur. */
4156
4157 if (all_reduced == 1 && bl->eliminable
4158 && maybe_eliminate_biv (bl, loop_start, end, 1,
4159 threshold, insn_count))
4160
4161 {
4162 /* ?? If we created a new test to bypass the loop entirely,
4163 or otherwise drop straight in, based on this test, then
4164 we might want to rewrite it also. This way some later
4165 pass has more hope of removing the initialization of this
4166 biv entirely. */
4167
4168 /* If final_value != 0, then the biv may be used after loop end
4169 and we must emit an insn to set it just in case.
4170
4171 Reversed bivs already have an insn after the loop setting their
4172 value, so we don't need another one. We can't calculate the
4173 proper final value for such a biv here anyways. */
4174 if (final_value != 0 && ! bl->reversed)
4175 {
4176 rtx insert_before;
4177
4178 /* If the loop has multiple exits, emit the insn before the
4179 loop to ensure that it will always be executed no matter
4180 how the loop exits. Otherwise, emit the insn after the
4181 loop, since this is slightly more efficient. */
4182 if (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
4183 insert_before = loop_start;
4184 else
4185 insert_before = end_insert_before;
4186
4187 emit_insn_before (gen_move_insn (bl->biv->dest_reg, final_value),
4188 end_insert_before);
4189 }
4190
4191 #if 0
4192 /* Delete all of the instructions inside the loop which set
4193 the biv, as they are all dead. If is safe to delete them,
4194 because an insn setting a biv will never be part of a libcall. */
4195 /* However, deleting them will invalidate the regno_last_uid info,
4196 so keeping them around is more convenient. Final_biv_value
4197 will only succeed when there are multiple exits if the biv
4198 is dead at each exit, hence it does not matter that the original
4199 insn remains, because it is dead anyways. */
4200 for (v = bl->biv; v; v = v->next_iv)
4201 delete_insn (v->insn);
4202 #endif
4203
4204 if (loop_dump_stream)
4205 fprintf (loop_dump_stream, "Reg %d: biv eliminated\n",
4206 bl->regno);
4207 }
4208 }
4209
4210 /* Go through all the instructions in the loop, making all the
4211 register substitutions scheduled in REG_MAP. */
4212
4213 for (p = loop_start; p != end; p = NEXT_INSN (p))
4214 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4215 || GET_CODE (p) == CALL_INSN)
4216 {
4217 replace_regs (PATTERN (p), reg_map, max_reg_before_loop, 0);
4218 replace_regs (REG_NOTES (p), reg_map, max_reg_before_loop, 0);
4219 INSN_CODE (p) = -1;
4220 }
4221
4222 /* Unroll loops from within strength reduction so that we can use the
4223 induction variable information that strength_reduce has already
4224 collected. */
4225
4226 if (flag_unroll_loops)
4227 unroll_loop (loop_end, insn_count, loop_start, end_insert_before, 1);
4228
4229 if (loop_dump_stream)
4230 fprintf (loop_dump_stream, "\n");
4231 }
4232 \f
4233 /* Return 1 if X is a valid source for an initial value (or as value being
4234 compared against in an initial test).
4235
4236 X must be either a register or constant and must not be clobbered between
4237 the current insn and the start of the loop.
4238
4239 INSN is the insn containing X. */
4240
4241 static int
4242 valid_initial_value_p (x, insn, call_seen, loop_start)
4243 rtx x;
4244 rtx insn;
4245 int call_seen;
4246 rtx loop_start;
4247 {
4248 if (CONSTANT_P (x))
4249 return 1;
4250
4251 /* Only consider pseudos we know about initialized in insns whose luids
4252 we know. */
4253 if (GET_CODE (x) != REG
4254 || REGNO (x) >= max_reg_before_loop)
4255 return 0;
4256
4257 /* Don't use call-clobbered registers across a call which clobbers it. On
4258 some machines, don't use any hard registers at all. */
4259 if (REGNO (x) < FIRST_PSEUDO_REGISTER
4260 #ifndef SMALL_REGISTER_CLASSES
4261 && call_used_regs[REGNO (x)] && call_seen
4262 #endif
4263 )
4264 return 0;
4265
4266 /* Don't use registers that have been clobbered before the start of the
4267 loop. */
4268 if (reg_set_between_p (x, insn, loop_start))
4269 return 0;
4270
4271 return 1;
4272 }
4273 \f
4274 /* Scan X for memory refs and check each memory address
4275 as a possible giv. INSN is the insn whose pattern X comes from.
4276 NOT_EVERY_ITERATION is 1 if the insn might not be executed during
4277 every loop iteration. */
4278
4279 static void
4280 find_mem_givs (x, insn, not_every_iteration, loop_start, loop_end)
4281 rtx x;
4282 rtx insn;
4283 int not_every_iteration;
4284 rtx loop_start, loop_end;
4285 {
4286 register int i, j;
4287 register enum rtx_code code;
4288 register char *fmt;
4289
4290 if (x == 0)
4291 return;
4292
4293 code = GET_CODE (x);
4294 switch (code)
4295 {
4296 case REG:
4297 case CONST_INT:
4298 case CONST:
4299 case CONST_DOUBLE:
4300 case SYMBOL_REF:
4301 case LABEL_REF:
4302 case PC:
4303 case CC0:
4304 case ADDR_VEC:
4305 case ADDR_DIFF_VEC:
4306 case USE:
4307 case CLOBBER:
4308 return;
4309
4310 case MEM:
4311 {
4312 rtx src_reg;
4313 rtx add_val;
4314 rtx mult_val;
4315 int benefit;
4316
4317 benefit = general_induction_var (XEXP (x, 0),
4318 &src_reg, &add_val, &mult_val);
4319
4320 /* Don't make a DEST_ADDR giv with mult_val == 1 && add_val == 0.
4321 Such a giv isn't useful. */
4322 if (benefit > 0 && (mult_val != const1_rtx || add_val != const0_rtx))
4323 {
4324 /* Found one; record it. */
4325 struct induction *v
4326 = (struct induction *) oballoc (sizeof (struct induction));
4327
4328 record_giv (v, insn, src_reg, addr_placeholder, mult_val,
4329 add_val, benefit, DEST_ADDR, not_every_iteration,
4330 &XEXP (x, 0), loop_start, loop_end);
4331
4332 v->mem_mode = GET_MODE (x);
4333 }
4334 return;
4335 }
4336 }
4337
4338 /* Recursively scan the subexpressions for other mem refs. */
4339
4340 fmt = GET_RTX_FORMAT (code);
4341 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4342 if (fmt[i] == 'e')
4343 find_mem_givs (XEXP (x, i), insn, not_every_iteration, loop_start,
4344 loop_end);
4345 else if (fmt[i] == 'E')
4346 for (j = 0; j < XVECLEN (x, i); j++)
4347 find_mem_givs (XVECEXP (x, i, j), insn, not_every_iteration,
4348 loop_start, loop_end);
4349 }
4350 \f
4351 /* Fill in the data about one biv update.
4352 V is the `struct induction' in which we record the biv. (It is
4353 allocated by the caller, with alloca.)
4354 INSN is the insn that sets it.
4355 DEST_REG is the biv's reg.
4356
4357 MULT_VAL is const1_rtx if the biv is being incremented here, in which case
4358 INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
4359 being set to INC_VAL.
4360
4361 NOT_EVERY_ITERATION is nonzero if this biv update is not know to be
4362 executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update
4363 can be executed more than once per iteration. If MAYBE_MULTIPLE
4364 and NOT_EVERY_ITERATION are both zero, we know that the biv update is
4365 executed exactly once per iteration. */
4366
4367 static void
4368 record_biv (v, insn, dest_reg, inc_val, mult_val,
4369 not_every_iteration, maybe_multiple)
4370 struct induction *v;
4371 rtx insn;
4372 rtx dest_reg;
4373 rtx inc_val;
4374 rtx mult_val;
4375 int not_every_iteration;
4376 int maybe_multiple;
4377 {
4378 struct iv_class *bl;
4379
4380 v->insn = insn;
4381 v->src_reg = dest_reg;
4382 v->dest_reg = dest_reg;
4383 v->mult_val = mult_val;
4384 v->add_val = inc_val;
4385 v->mode = GET_MODE (dest_reg);
4386 v->always_computable = ! not_every_iteration;
4387 v->always_executed = ! not_every_iteration;
4388 v->maybe_multiple = maybe_multiple;
4389
4390 /* Add this to the reg's iv_class, creating a class
4391 if this is the first incrementation of the reg. */
4392
4393 bl = reg_biv_class[REGNO (dest_reg)];
4394 if (bl == 0)
4395 {
4396 /* Create and initialize new iv_class. */
4397
4398 bl = (struct iv_class *) oballoc (sizeof (struct iv_class));
4399
4400 bl->regno = REGNO (dest_reg);
4401 bl->biv = 0;
4402 bl->giv = 0;
4403 bl->biv_count = 0;
4404 bl->giv_count = 0;
4405
4406 /* Set initial value to the reg itself. */
4407 bl->initial_value = dest_reg;
4408 /* We haven't seen the initializing insn yet */
4409 bl->init_insn = 0;
4410 bl->init_set = 0;
4411 bl->initial_test = 0;
4412 bl->incremented = 0;
4413 bl->eliminable = 0;
4414 bl->nonneg = 0;
4415 bl->reversed = 0;
4416 bl->total_benefit = 0;
4417
4418 /* Add this class to loop_iv_list. */
4419 bl->next = loop_iv_list;
4420 loop_iv_list = bl;
4421
4422 /* Put it in the array of biv register classes. */
4423 reg_biv_class[REGNO (dest_reg)] = bl;
4424 }
4425
4426 /* Update IV_CLASS entry for this biv. */
4427 v->next_iv = bl->biv;
4428 bl->biv = v;
4429 bl->biv_count++;
4430 if (mult_val == const1_rtx)
4431 bl->incremented = 1;
4432
4433 if (loop_dump_stream)
4434 {
4435 fprintf (loop_dump_stream,
4436 "Insn %d: possible biv, reg %d,",
4437 INSN_UID (insn), REGNO (dest_reg));
4438 if (GET_CODE (inc_val) == CONST_INT)
4439 fprintf (loop_dump_stream, " const = %d\n",
4440 INTVAL (inc_val));
4441 else
4442 {
4443 fprintf (loop_dump_stream, " const = ");
4444 print_rtl (loop_dump_stream, inc_val);
4445 fprintf (loop_dump_stream, "\n");
4446 }
4447 }
4448 }
4449 \f
4450 /* Fill in the data about one giv.
4451 V is the `struct induction' in which we record the giv. (It is
4452 allocated by the caller, with alloca.)
4453 INSN is the insn that sets it.
4454 BENEFIT estimates the savings from deleting this insn.
4455 TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
4456 into a register or is used as a memory address.
4457
4458 SRC_REG is the biv reg which the giv is computed from.
4459 DEST_REG is the giv's reg (if the giv is stored in a reg).
4460 MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
4461 LOCATION points to the place where this giv's value appears in INSN. */
4462
4463 static void
4464 record_giv (v, insn, src_reg, dest_reg, mult_val, add_val, benefit,
4465 type, not_every_iteration, location, loop_start, loop_end)
4466 struct induction *v;
4467 rtx insn;
4468 rtx src_reg;
4469 rtx dest_reg;
4470 rtx mult_val, add_val;
4471 int benefit;
4472 enum g_types type;
4473 int not_every_iteration;
4474 rtx *location;
4475 rtx loop_start, loop_end;
4476 {
4477 struct induction *b;
4478 struct iv_class *bl;
4479 rtx set = single_set (insn);
4480 rtx p;
4481
4482 v->insn = insn;
4483 v->src_reg = src_reg;
4484 v->giv_type = type;
4485 v->dest_reg = dest_reg;
4486 v->mult_val = mult_val;
4487 v->add_val = add_val;
4488 v->benefit = benefit;
4489 v->location = location;
4490 v->cant_derive = 0;
4491 v->combined_with = 0;
4492 v->maybe_multiple = 0;
4493 v->maybe_dead = 0;
4494 v->derive_adjustment = 0;
4495 v->same = 0;
4496 v->ignore = 0;
4497 v->new_reg = 0;
4498 v->final_value = 0;
4499 v->same_insn = 0;
4500 v->auto_inc_opt = 0;
4501
4502 /* The v->always_computable field is used in update_giv_derive, to
4503 determine whether a giv can be used to derive another giv. For a
4504 DEST_REG giv, INSN computes a new value for the giv, so its value
4505 isn't computable if INSN insn't executed every iteration.
4506 However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
4507 it does not compute a new value. Hence the value is always computable
4508 regardless of whether INSN is executed each iteration. */
4509
4510 if (type == DEST_ADDR)
4511 v->always_computable = 1;
4512 else
4513 v->always_computable = ! not_every_iteration;
4514
4515 v->always_executed = ! not_every_iteration;
4516
4517 if (type == DEST_ADDR)
4518 {
4519 v->mode = GET_MODE (*location);
4520 v->lifetime = 1;
4521 v->times_used = 1;
4522 }
4523 else /* type == DEST_REG */
4524 {
4525 v->mode = GET_MODE (SET_DEST (set));
4526
4527 v->lifetime = (uid_luid[regno_last_uid[REGNO (dest_reg)]]
4528 - uid_luid[regno_first_uid[REGNO (dest_reg)]]);
4529
4530 v->times_used = n_times_used[REGNO (dest_reg)];
4531
4532 /* If the lifetime is zero, it means that this register is
4533 really a dead store. So mark this as a giv that can be
4534 ignored. This will not prevent the biv from being eliminated. */
4535 if (v->lifetime == 0)
4536 v->ignore = 1;
4537
4538 reg_iv_type[REGNO (dest_reg)] = GENERAL_INDUCT;
4539 reg_iv_info[REGNO (dest_reg)] = v;
4540 }
4541
4542 /* Add the giv to the class of givs computed from one biv. */
4543
4544 bl = reg_biv_class[REGNO (src_reg)];
4545 if (bl)
4546 {
4547 v->next_iv = bl->giv;
4548 bl->giv = v;
4549 /* Don't count DEST_ADDR. This is supposed to count the number of
4550 insns that calculate givs. */
4551 if (type == DEST_REG)
4552 bl->giv_count++;
4553 bl->total_benefit += benefit;
4554 }
4555 else
4556 /* Fatal error, biv missing for this giv? */
4557 abort ();
4558
4559 if (type == DEST_ADDR)
4560 v->replaceable = 1;
4561 else
4562 {
4563 /* The giv can be replaced outright by the reduced register only if all
4564 of the following conditions are true:
4565 - the insn that sets the giv is always executed on any iteration
4566 on which the giv is used at all
4567 (there are two ways to deduce this:
4568 either the insn is executed on every iteration,
4569 or all uses follow that insn in the same basic block),
4570 - the giv is not used outside the loop
4571 - no assignments to the biv occur during the giv's lifetime. */
4572
4573 if (regno_first_uid[REGNO (dest_reg)] == INSN_UID (insn)
4574 /* Previous line always fails if INSN was moved by loop opt. */
4575 && uid_luid[regno_last_uid[REGNO (dest_reg)]] < INSN_LUID (loop_end)
4576 && (! not_every_iteration
4577 || last_use_this_basic_block (dest_reg, insn)))
4578 {
4579 /* Now check that there are no assignments to the biv within the
4580 giv's lifetime. This requires two separate checks. */
4581
4582 /* Check each biv update, and fail if any are between the first
4583 and last use of the giv.
4584
4585 If this loop contains an inner loop that was unrolled, then
4586 the insn modifying the biv may have been emitted by the loop
4587 unrolling code, and hence does not have a valid luid. Just
4588 mark the biv as not replaceable in this case. It is not very
4589 useful as a biv, because it is used in two different loops.
4590 It is very unlikely that we would be able to optimize the giv
4591 using this biv anyways. */
4592
4593 v->replaceable = 1;
4594 for (b = bl->biv; b; b = b->next_iv)
4595 {
4596 if (INSN_UID (b->insn) >= max_uid_for_loop
4597 || ((uid_luid[INSN_UID (b->insn)]
4598 >= uid_luid[regno_first_uid[REGNO (dest_reg)]])
4599 && (uid_luid[INSN_UID (b->insn)]
4600 <= uid_luid[regno_last_uid[REGNO (dest_reg)]])))
4601 {
4602 v->replaceable = 0;
4603 v->not_replaceable = 1;
4604 break;
4605 }
4606 }
4607
4608 /* If there are any backwards branches that go from after the
4609 biv update to before it, then this giv is not replaceable. */
4610 if (v->replaceable)
4611 for (b = bl->biv; b; b = b->next_iv)
4612 if (back_branch_in_range_p (b->insn, loop_start, loop_end))
4613 {
4614 v->replaceable = 0;
4615 v->not_replaceable = 1;
4616 break;
4617 }
4618 }
4619 else
4620 {
4621 /* May still be replaceable, we don't have enough info here to
4622 decide. */
4623 v->replaceable = 0;
4624 v->not_replaceable = 0;
4625 }
4626 }
4627
4628 if (loop_dump_stream)
4629 {
4630 if (type == DEST_REG)
4631 fprintf (loop_dump_stream, "Insn %d: giv reg %d",
4632 INSN_UID (insn), REGNO (dest_reg));
4633 else
4634 fprintf (loop_dump_stream, "Insn %d: dest address",
4635 INSN_UID (insn));
4636
4637 fprintf (loop_dump_stream, " src reg %d benefit %d",
4638 REGNO (src_reg), v->benefit);
4639 fprintf (loop_dump_stream, " used %d lifetime %d",
4640 v->times_used, v->lifetime);
4641
4642 if (v->replaceable)
4643 fprintf (loop_dump_stream, " replaceable");
4644
4645 if (GET_CODE (mult_val) == CONST_INT)
4646 fprintf (loop_dump_stream, " mult %d",
4647 INTVAL (mult_val));
4648 else
4649 {
4650 fprintf (loop_dump_stream, " mult ");
4651 print_rtl (loop_dump_stream, mult_val);
4652 }
4653
4654 if (GET_CODE (add_val) == CONST_INT)
4655 fprintf (loop_dump_stream, " add %d",
4656 INTVAL (add_val));
4657 else
4658 {
4659 fprintf (loop_dump_stream, " add ");
4660 print_rtl (loop_dump_stream, add_val);
4661 }
4662 }
4663
4664 if (loop_dump_stream)
4665 fprintf (loop_dump_stream, "\n");
4666
4667 }
4668
4669
4670 /* All this does is determine whether a giv can be made replaceable because
4671 its final value can be calculated. This code can not be part of record_giv
4672 above, because final_giv_value requires that the number of loop iterations
4673 be known, and that can not be accurately calculated until after all givs
4674 have been identified. */
4675
4676 static void
4677 check_final_value (v, loop_start, loop_end)
4678 struct induction *v;
4679 rtx loop_start, loop_end;
4680 {
4681 struct iv_class *bl;
4682 rtx final_value = 0;
4683
4684 bl = reg_biv_class[REGNO (v->src_reg)];
4685
4686 /* DEST_ADDR givs will never reach here, because they are always marked
4687 replaceable above in record_giv. */
4688
4689 /* The giv can be replaced outright by the reduced register only if all
4690 of the following conditions are true:
4691 - the insn that sets the giv is always executed on any iteration
4692 on which the giv is used at all
4693 (there are two ways to deduce this:
4694 either the insn is executed on every iteration,
4695 or all uses follow that insn in the same basic block),
4696 - its final value can be calculated (this condition is different
4697 than the one above in record_giv)
4698 - no assignments to the biv occur during the giv's lifetime. */
4699
4700 #if 0
4701 /* This is only called now when replaceable is known to be false. */
4702 /* Clear replaceable, so that it won't confuse final_giv_value. */
4703 v->replaceable = 0;
4704 #endif
4705
4706 if ((final_value = final_giv_value (v, loop_start, loop_end))
4707 && (v->always_computable || last_use_this_basic_block (v->dest_reg, v->insn)))
4708 {
4709 int biv_increment_seen = 0;
4710 rtx p = v->insn;
4711 rtx last_giv_use;
4712
4713 v->replaceable = 1;
4714
4715 /* When trying to determine whether or not a biv increment occurs
4716 during the lifetime of the giv, we can ignore uses of the variable
4717 outside the loop because final_value is true. Hence we can not
4718 use regno_last_uid and regno_first_uid as above in record_giv. */
4719
4720 /* Search the loop to determine whether any assignments to the
4721 biv occur during the giv's lifetime. Start with the insn
4722 that sets the giv, and search around the loop until we come
4723 back to that insn again.
4724
4725 Also fail if there is a jump within the giv's lifetime that jumps
4726 to somewhere outside the lifetime but still within the loop. This
4727 catches spaghetti code where the execution order is not linear, and
4728 hence the above test fails. Here we assume that the giv lifetime
4729 does not extend from one iteration of the loop to the next, so as
4730 to make the test easier. Since the lifetime isn't known yet,
4731 this requires two loops. See also record_giv above. */
4732
4733 last_giv_use = v->insn;
4734
4735 while (1)
4736 {
4737 p = NEXT_INSN (p);
4738 if (p == loop_end)
4739 p = NEXT_INSN (loop_start);
4740 if (p == v->insn)
4741 break;
4742
4743 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4744 || GET_CODE (p) == CALL_INSN)
4745 {
4746 if (biv_increment_seen)
4747 {
4748 if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
4749 {
4750 v->replaceable = 0;
4751 v->not_replaceable = 1;
4752 break;
4753 }
4754 }
4755 else if (GET_CODE (PATTERN (p)) == SET
4756 && SET_DEST (PATTERN (p)) == v->src_reg)
4757 biv_increment_seen = 1;
4758 else if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
4759 last_giv_use = p;
4760 }
4761 }
4762
4763 /* Now that the lifetime of the giv is known, check for branches
4764 from within the lifetime to outside the lifetime if it is still
4765 replaceable. */
4766
4767 if (v->replaceable)
4768 {
4769 p = v->insn;
4770 while (1)
4771 {
4772 p = NEXT_INSN (p);
4773 if (p == loop_end)
4774 p = NEXT_INSN (loop_start);
4775 if (p == last_giv_use)
4776 break;
4777
4778 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
4779 && LABEL_NAME (JUMP_LABEL (p))
4780 && ((INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (v->insn)
4781 && INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (loop_start))
4782 || (INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (last_giv_use)
4783 && INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (loop_end))))
4784 {
4785 v->replaceable = 0;
4786 v->not_replaceable = 1;
4787
4788 if (loop_dump_stream)
4789 fprintf (loop_dump_stream,
4790 "Found branch outside giv lifetime.\n");
4791
4792 break;
4793 }
4794 }
4795 }
4796
4797 /* If it is replaceable, then save the final value. */
4798 if (v->replaceable)
4799 v->final_value = final_value;
4800 }
4801
4802 if (loop_dump_stream && v->replaceable)
4803 fprintf (loop_dump_stream, "Insn %d: giv reg %d final_value replaceable\n",
4804 INSN_UID (v->insn), REGNO (v->dest_reg));
4805 }
4806 \f
4807 /* Update the status of whether a giv can derive other givs.
4808
4809 We need to do something special if there is or may be an update to the biv
4810 between the time the giv is defined and the time it is used to derive
4811 another giv.
4812
4813 In addition, a giv that is only conditionally set is not allowed to
4814 derive another giv once a label has been passed.
4815
4816 The cases we look at are when a label or an update to a biv is passed. */
4817
4818 static void
4819 update_giv_derive (p)
4820 rtx p;
4821 {
4822 struct iv_class *bl;
4823 struct induction *biv, *giv;
4824 rtx tem;
4825 int dummy;
4826
4827 /* Search all IV classes, then all bivs, and finally all givs.
4828
4829 There are three cases we are concerned with. First we have the situation
4830 of a giv that is only updated conditionally. In that case, it may not
4831 derive any givs after a label is passed.
4832
4833 The second case is when a biv update occurs, or may occur, after the
4834 definition of a giv. For certain biv updates (see below) that are
4835 known to occur between the giv definition and use, we can adjust the
4836 giv definition. For others, or when the biv update is conditional,
4837 we must prevent the giv from deriving any other givs. There are two
4838 sub-cases within this case.
4839
4840 If this is a label, we are concerned with any biv update that is done
4841 conditionally, since it may be done after the giv is defined followed by
4842 a branch here (actually, we need to pass both a jump and a label, but
4843 this extra tracking doesn't seem worth it).
4844
4845 If this is a jump, we are concerned about any biv update that may be
4846 executed multiple times. We are actually only concerned about
4847 backward jumps, but it is probably not worth performing the test
4848 on the jump again here.
4849
4850 If this is a biv update, we must adjust the giv status to show that a
4851 subsequent biv update was performed. If this adjustment cannot be done,
4852 the giv cannot derive further givs. */
4853
4854 for (bl = loop_iv_list; bl; bl = bl->next)
4855 for (biv = bl->biv; biv; biv = biv->next_iv)
4856 if (GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN
4857 || biv->insn == p)
4858 {
4859 for (giv = bl->giv; giv; giv = giv->next_iv)
4860 {
4861 /* If cant_derive is already true, there is no point in
4862 checking all of these conditions again. */
4863 if (giv->cant_derive)
4864 continue;
4865
4866 /* If this giv is conditionally set and we have passed a label,
4867 it cannot derive anything. */
4868 if (GET_CODE (p) == CODE_LABEL && ! giv->always_computable)
4869 giv->cant_derive = 1;
4870
4871 /* Skip givs that have mult_val == 0, since
4872 they are really invariants. Also skip those that are
4873 replaceable, since we know their lifetime doesn't contain
4874 any biv update. */
4875 else if (giv->mult_val == const0_rtx || giv->replaceable)
4876 continue;
4877
4878 /* The only way we can allow this giv to derive another
4879 is if this is a biv increment and we can form the product
4880 of biv->add_val and giv->mult_val. In this case, we will
4881 be able to compute a compensation. */
4882 else if (biv->insn == p)
4883 {
4884 tem = 0;
4885
4886 if (biv->mult_val == const1_rtx)
4887 tem = simplify_giv_expr (gen_rtx (MULT, giv->mode,
4888 biv->add_val,
4889 giv->mult_val),
4890 &dummy);
4891
4892 if (tem && giv->derive_adjustment)
4893 tem = simplify_giv_expr (gen_rtx (PLUS, giv->mode, tem,
4894 giv->derive_adjustment),
4895 &dummy);
4896 if (tem)
4897 giv->derive_adjustment = tem;
4898 else
4899 giv->cant_derive = 1;
4900 }
4901 else if ((GET_CODE (p) == CODE_LABEL && ! biv->always_computable)
4902 || (GET_CODE (p) == JUMP_INSN && biv->maybe_multiple))
4903 giv->cant_derive = 1;
4904 }
4905 }
4906 }
4907 \f
4908 /* Check whether an insn is an increment legitimate for a basic induction var.
4909 X is the source of insn P, or a part of it.
4910 MODE is the mode in which X should be interpreted.
4911
4912 DEST_REG is the putative biv, also the destination of the insn.
4913 We accept patterns of these forms:
4914 REG = REG + INVARIANT (includes REG = REG - CONSTANT)
4915 REG = INVARIANT + REG
4916
4917 If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
4918 and store the additive term into *INC_VAL.
4919
4920 If X is an assignment of an invariant into DEST_REG, we set
4921 *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
4922
4923 We also want to detect a BIV when it corresponds to a variable
4924 whose mode was promoted via PROMOTED_MODE. In that case, an increment
4925 of the variable may be a PLUS that adds a SUBREG of that variable to
4926 an invariant and then sign- or zero-extends the result of the PLUS
4927 into the variable.
4928
4929 Most GIVs in such cases will be in the promoted mode, since that is the
4930 probably the natural computation mode (and almost certainly the mode
4931 used for addresses) on the machine. So we view the pseudo-reg containing
4932 the variable as the BIV, as if it were simply incremented.
4933
4934 Note that treating the entire pseudo as a BIV will result in making
4935 simple increments to any GIVs based on it. However, if the variable
4936 overflows in its declared mode but not its promoted mode, the result will
4937 be incorrect. This is acceptable if the variable is signed, since
4938 overflows in such cases are undefined, but not if it is unsigned, since
4939 those overflows are defined. So we only check for SIGN_EXTEND and
4940 not ZERO_EXTEND.
4941
4942 If we cannot find a biv, we return 0. */
4943
4944 static int
4945 basic_induction_var (x, mode, dest_reg, p, inc_val, mult_val)
4946 register rtx x;
4947 enum machine_mode mode;
4948 rtx p;
4949 rtx dest_reg;
4950 rtx *inc_val;
4951 rtx *mult_val;
4952 {
4953 register enum rtx_code code;
4954 rtx arg;
4955 rtx insn, set = 0;
4956
4957 code = GET_CODE (x);
4958 switch (code)
4959 {
4960 case PLUS:
4961 if (XEXP (x, 0) == dest_reg
4962 || (GET_CODE (XEXP (x, 0)) == SUBREG
4963 && SUBREG_PROMOTED_VAR_P (XEXP (x, 0))
4964 && SUBREG_REG (XEXP (x, 0)) == dest_reg))
4965 arg = XEXP (x, 1);
4966 else if (XEXP (x, 1) == dest_reg
4967 || (GET_CODE (XEXP (x, 1)) == SUBREG
4968 && SUBREG_PROMOTED_VAR_P (XEXP (x, 1))
4969 && SUBREG_REG (XEXP (x, 1)) == dest_reg))
4970 arg = XEXP (x, 0);
4971 else
4972 return 0;
4973
4974 if (invariant_p (arg) != 1)
4975 return 0;
4976
4977 *inc_val = convert_modes (GET_MODE (dest_reg), GET_MODE (x), arg, 0);
4978 *mult_val = const1_rtx;
4979 return 1;
4980
4981 case SUBREG:
4982 /* If this is a SUBREG for a promoted variable, check the inner
4983 value. */
4984 if (SUBREG_PROMOTED_VAR_P (x))
4985 return basic_induction_var (SUBREG_REG (x), GET_MODE (SUBREG_REG (x)),
4986 dest_reg, p, inc_val, mult_val);
4987
4988 case REG:
4989 /* If this register is assigned in the previous insn, look at its
4990 source, but don't go outside the loop or past a label. */
4991
4992 for (insn = PREV_INSN (p);
4993 (insn && GET_CODE (insn) == NOTE
4994 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
4995 insn = PREV_INSN (insn))
4996 ;
4997
4998 if (insn)
4999 set = single_set (insn);
5000
5001 if (set != 0
5002 && (SET_DEST (set) == x
5003 || (GET_CODE (SET_DEST (set)) == SUBREG
5004 && (GET_MODE_SIZE (GET_MODE (SET_DEST (set)))
5005 <= UNITS_PER_WORD)
5006 && SUBREG_REG (SET_DEST (set)) == x)))
5007 return basic_induction_var (SET_SRC (set),
5008 (GET_MODE (SET_SRC (set)) == VOIDmode
5009 ? GET_MODE (x)
5010 : GET_MODE (SET_SRC (set))),
5011 dest_reg, insn,
5012 inc_val, mult_val);
5013 /* ... fall through ... */
5014
5015 /* Can accept constant setting of biv only when inside inner most loop.
5016 Otherwise, a biv of an inner loop may be incorrectly recognized
5017 as a biv of the outer loop,
5018 causing code to be moved INTO the inner loop. */
5019 case MEM:
5020 if (invariant_p (x) != 1)
5021 return 0;
5022 case CONST_INT:
5023 case SYMBOL_REF:
5024 case CONST:
5025 if (loops_enclosed == 1)
5026 {
5027 /* Possible bug here? Perhaps we don't know the mode of X. */
5028 *inc_val = convert_modes (GET_MODE (dest_reg), mode, x, 0);
5029 *mult_val = const0_rtx;
5030 return 1;
5031 }
5032 else
5033 return 0;
5034
5035 case SIGN_EXTEND:
5036 return basic_induction_var (XEXP (x, 0), GET_MODE (XEXP (x, 0)),
5037 dest_reg, p, inc_val, mult_val);
5038 case ASHIFTRT:
5039 /* Similar, since this can be a sign extension. */
5040 for (insn = PREV_INSN (p);
5041 (insn && GET_CODE (insn) == NOTE
5042 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
5043 insn = PREV_INSN (insn))
5044 ;
5045
5046 if (insn)
5047 set = single_set (insn);
5048
5049 if (set && SET_DEST (set) == XEXP (x, 0)
5050 && GET_CODE (XEXP (x, 1)) == CONST_INT
5051 && INTVAL (XEXP (x, 1)) >= 0
5052 && GET_CODE (SET_SRC (set)) == ASHIFT
5053 && XEXP (x, 1) == XEXP (SET_SRC (set), 1))
5054 return basic_induction_var (XEXP (SET_SRC (set), 0),
5055 GET_MODE (XEXP (x, 0)),
5056 dest_reg, insn, inc_val, mult_val);
5057 return 0;
5058
5059 default:
5060 return 0;
5061 }
5062 }
5063 \f
5064 /* A general induction variable (giv) is any quantity that is a linear
5065 function of a basic induction variable,
5066 i.e. giv = biv * mult_val + add_val.
5067 The coefficients can be any loop invariant quantity.
5068 A giv need not be computed directly from the biv;
5069 it can be computed by way of other givs. */
5070
5071 /* Determine whether X computes a giv.
5072 If it does, return a nonzero value
5073 which is the benefit from eliminating the computation of X;
5074 set *SRC_REG to the register of the biv that it is computed from;
5075 set *ADD_VAL and *MULT_VAL to the coefficients,
5076 such that the value of X is biv * mult + add; */
5077
5078 static int
5079 general_induction_var (x, src_reg, add_val, mult_val)
5080 rtx x;
5081 rtx *src_reg;
5082 rtx *add_val;
5083 rtx *mult_val;
5084 {
5085 rtx orig_x = x;
5086 int benefit = 0;
5087 char *storage;
5088
5089 /* If this is an invariant, forget it, it isn't a giv. */
5090 if (invariant_p (x) == 1)
5091 return 0;
5092
5093 /* See if the expression could be a giv and get its form.
5094 Mark our place on the obstack in case we don't find a giv. */
5095 storage = (char *) oballoc (0);
5096 x = simplify_giv_expr (x, &benefit);
5097 if (x == 0)
5098 {
5099 obfree (storage);
5100 return 0;
5101 }
5102
5103 switch (GET_CODE (x))
5104 {
5105 case USE:
5106 case CONST_INT:
5107 /* Since this is now an invariant and wasn't before, it must be a giv
5108 with MULT_VAL == 0. It doesn't matter which BIV we associate this
5109 with. */
5110 *src_reg = loop_iv_list->biv->dest_reg;
5111 *mult_val = const0_rtx;
5112 *add_val = x;
5113 break;
5114
5115 case REG:
5116 /* This is equivalent to a BIV. */
5117 *src_reg = x;
5118 *mult_val = const1_rtx;
5119 *add_val = const0_rtx;
5120 break;
5121
5122 case PLUS:
5123 /* Either (plus (biv) (invar)) or
5124 (plus (mult (biv) (invar_1)) (invar_2)). */
5125 if (GET_CODE (XEXP (x, 0)) == MULT)
5126 {
5127 *src_reg = XEXP (XEXP (x, 0), 0);
5128 *mult_val = XEXP (XEXP (x, 0), 1);
5129 }
5130 else
5131 {
5132 *src_reg = XEXP (x, 0);
5133 *mult_val = const1_rtx;
5134 }
5135 *add_val = XEXP (x, 1);
5136 break;
5137
5138 case MULT:
5139 /* ADD_VAL is zero. */
5140 *src_reg = XEXP (x, 0);
5141 *mult_val = XEXP (x, 1);
5142 *add_val = const0_rtx;
5143 break;
5144
5145 default:
5146 abort ();
5147 }
5148
5149 /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
5150 unless they are CONST_INT). */
5151 if (GET_CODE (*add_val) == USE)
5152 *add_val = XEXP (*add_val, 0);
5153 if (GET_CODE (*mult_val) == USE)
5154 *mult_val = XEXP (*mult_val, 0);
5155
5156 benefit += rtx_cost (orig_x, SET);
5157
5158 /* Always return some benefit if this is a giv so it will be detected
5159 as such. This allows elimination of bivs that might otherwise
5160 not be eliminated. */
5161 return benefit == 0 ? 1 : benefit;
5162 }
5163 \f
5164 /* Given an expression, X, try to form it as a linear function of a biv.
5165 We will canonicalize it to be of the form
5166 (plus (mult (BIV) (invar_1))
5167 (invar_2))
5168 with possible degeneracies.
5169
5170 The invariant expressions must each be of a form that can be used as a
5171 machine operand. We surround then with a USE rtx (a hack, but localized
5172 and certainly unambiguous!) if not a CONST_INT for simplicity in this
5173 routine; it is the caller's responsibility to strip them.
5174
5175 If no such canonicalization is possible (i.e., two biv's are used or an
5176 expression that is neither invariant nor a biv or giv), this routine
5177 returns 0.
5178
5179 For a non-zero return, the result will have a code of CONST_INT, USE,
5180 REG (for a BIV), PLUS, or MULT. No other codes will occur.
5181
5182 *BENEFIT will be incremented by the benefit of any sub-giv encountered. */
5183
5184 static rtx
5185 simplify_giv_expr (x, benefit)
5186 rtx x;
5187 int *benefit;
5188 {
5189 enum machine_mode mode = GET_MODE (x);
5190 rtx arg0, arg1;
5191 rtx tem;
5192
5193 /* If this is not an integer mode, or if we cannot do arithmetic in this
5194 mode, this can't be a giv. */
5195 if (mode != VOIDmode
5196 && (GET_MODE_CLASS (mode) != MODE_INT
5197 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT))
5198 return 0;
5199
5200 switch (GET_CODE (x))
5201 {
5202 case PLUS:
5203 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
5204 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
5205 if (arg0 == 0 || arg1 == 0)
5206 return 0;
5207
5208 /* Put constant last, CONST_INT last if both constant. */
5209 if ((GET_CODE (arg0) == USE
5210 || GET_CODE (arg0) == CONST_INT)
5211 && GET_CODE (arg1) != CONST_INT)
5212 tem = arg0, arg0 = arg1, arg1 = tem;
5213
5214 /* Handle addition of zero, then addition of an invariant. */
5215 if (arg1 == const0_rtx)
5216 return arg0;
5217 else if (GET_CODE (arg1) == CONST_INT || GET_CODE (arg1) == USE)
5218 switch (GET_CODE (arg0))
5219 {
5220 case CONST_INT:
5221 case USE:
5222 /* Both invariant. Only valid if sum is machine operand.
5223 First strip off possible USE on first operand. */
5224 if (GET_CODE (arg0) == USE)
5225 arg0 = XEXP (arg0, 0);
5226
5227 tem = 0;
5228 if (CONSTANT_P (arg0) && GET_CODE (arg1) == CONST_INT)
5229 {
5230 tem = plus_constant (arg0, INTVAL (arg1));
5231 if (GET_CODE (tem) != CONST_INT)
5232 tem = gen_rtx (USE, mode, tem);
5233 }
5234
5235 return tem;
5236
5237 case REG:
5238 case MULT:
5239 /* biv + invar or mult + invar. Return sum. */
5240 return gen_rtx (PLUS, mode, arg0, arg1);
5241
5242 case PLUS:
5243 /* (a + invar_1) + invar_2. Associate. */
5244 return simplify_giv_expr (gen_rtx (PLUS, mode,
5245 XEXP (arg0, 0),
5246 gen_rtx (PLUS, mode,
5247 XEXP (arg0, 1), arg1)),
5248 benefit);
5249
5250 default:
5251 abort ();
5252 }
5253
5254 /* Each argument must be either REG, PLUS, or MULT. Convert REG to
5255 MULT to reduce cases. */
5256 if (GET_CODE (arg0) == REG)
5257 arg0 = gen_rtx (MULT, mode, arg0, const1_rtx);
5258 if (GET_CODE (arg1) == REG)
5259 arg1 = gen_rtx (MULT, mode, arg1, const1_rtx);
5260
5261 /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
5262 Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
5263 Recurse to associate the second PLUS. */
5264 if (GET_CODE (arg1) == MULT)
5265 tem = arg0, arg0 = arg1, arg1 = tem;
5266
5267 if (GET_CODE (arg1) == PLUS)
5268 return simplify_giv_expr (gen_rtx (PLUS, mode,
5269 gen_rtx (PLUS, mode,
5270 arg0, XEXP (arg1, 0)),
5271 XEXP (arg1, 1)),
5272 benefit);
5273
5274 /* Now must have MULT + MULT. Distribute if same biv, else not giv. */
5275 if (GET_CODE (arg0) != MULT || GET_CODE (arg1) != MULT)
5276 abort ();
5277
5278 if (XEXP (arg0, 0) != XEXP (arg1, 0))
5279 return 0;
5280
5281 return simplify_giv_expr (gen_rtx (MULT, mode,
5282 XEXP (arg0, 0),
5283 gen_rtx (PLUS, mode,
5284 XEXP (arg0, 1),
5285 XEXP (arg1, 1))),
5286 benefit);
5287
5288 case MINUS:
5289 /* Handle "a - b" as "a + b * (-1)". */
5290 return simplify_giv_expr (gen_rtx (PLUS, mode,
5291 XEXP (x, 0),
5292 gen_rtx (MULT, mode,
5293 XEXP (x, 1), constm1_rtx)),
5294 benefit);
5295
5296 case MULT:
5297 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
5298 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
5299 if (arg0 == 0 || arg1 == 0)
5300 return 0;
5301
5302 /* Put constant last, CONST_INT last if both constant. */
5303 if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT)
5304 && GET_CODE (arg1) != CONST_INT)
5305 tem = arg0, arg0 = arg1, arg1 = tem;
5306
5307 /* If second argument is not now constant, not giv. */
5308 if (GET_CODE (arg1) != USE && GET_CODE (arg1) != CONST_INT)
5309 return 0;
5310
5311 /* Handle multiply by 0 or 1. */
5312 if (arg1 == const0_rtx)
5313 return const0_rtx;
5314
5315 else if (arg1 == const1_rtx)
5316 return arg0;
5317
5318 switch (GET_CODE (arg0))
5319 {
5320 case REG:
5321 /* biv * invar. Done. */
5322 return gen_rtx (MULT, mode, arg0, arg1);
5323
5324 case CONST_INT:
5325 /* Product of two constants. */
5326 return GEN_INT (INTVAL (arg0) * INTVAL (arg1));
5327
5328 case USE:
5329 /* invar * invar. Not giv. */
5330 return 0;
5331
5332 case MULT:
5333 /* (a * invar_1) * invar_2. Associate. */
5334 return simplify_giv_expr (gen_rtx (MULT, mode,
5335 XEXP (arg0, 0),
5336 gen_rtx (MULT, mode,
5337 XEXP (arg0, 1), arg1)),
5338 benefit);
5339
5340 case PLUS:
5341 /* (a + invar_1) * invar_2. Distribute. */
5342 return simplify_giv_expr (gen_rtx (PLUS, mode,
5343 gen_rtx (MULT, mode,
5344 XEXP (arg0, 0), arg1),
5345 gen_rtx (MULT, mode,
5346 XEXP (arg0, 1), arg1)),
5347 benefit);
5348
5349 default:
5350 abort ();
5351 }
5352
5353 case ASHIFT:
5354 /* Shift by constant is multiply by power of two. */
5355 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
5356 return 0;
5357
5358 return simplify_giv_expr (gen_rtx (MULT, mode,
5359 XEXP (x, 0),
5360 GEN_INT ((HOST_WIDE_INT) 1
5361 << INTVAL (XEXP (x, 1)))),
5362 benefit);
5363
5364 case NEG:
5365 /* "-a" is "a * (-1)" */
5366 return simplify_giv_expr (gen_rtx (MULT, mode, XEXP (x, 0), constm1_rtx),
5367 benefit);
5368
5369 case NOT:
5370 /* "~a" is "-a - 1". Silly, but easy. */
5371 return simplify_giv_expr (gen_rtx (MINUS, mode,
5372 gen_rtx (NEG, mode, XEXP (x, 0)),
5373 const1_rtx),
5374 benefit);
5375
5376 case USE:
5377 /* Already in proper form for invariant. */
5378 return x;
5379
5380 case REG:
5381 /* If this is a new register, we can't deal with it. */
5382 if (REGNO (x) >= max_reg_before_loop)
5383 return 0;
5384
5385 /* Check for biv or giv. */
5386 switch (reg_iv_type[REGNO (x)])
5387 {
5388 case BASIC_INDUCT:
5389 return x;
5390 case GENERAL_INDUCT:
5391 {
5392 struct induction *v = reg_iv_info[REGNO (x)];
5393
5394 /* Form expression from giv and add benefit. Ensure this giv
5395 can derive another and subtract any needed adjustment if so. */
5396 *benefit += v->benefit;
5397 if (v->cant_derive)
5398 return 0;
5399
5400 tem = gen_rtx (PLUS, mode, gen_rtx (MULT, mode,
5401 v->src_reg, v->mult_val),
5402 v->add_val);
5403 if (v->derive_adjustment)
5404 tem = gen_rtx (MINUS, mode, tem, v->derive_adjustment);
5405 return simplify_giv_expr (tem, benefit);
5406 }
5407 }
5408
5409 /* Fall through to general case. */
5410 default:
5411 /* If invariant, return as USE (unless CONST_INT).
5412 Otherwise, not giv. */
5413 if (GET_CODE (x) == USE)
5414 x = XEXP (x, 0);
5415
5416 if (invariant_p (x) == 1)
5417 {
5418 if (GET_CODE (x) == CONST_INT)
5419 return x;
5420 else
5421 return gen_rtx (USE, mode, x);
5422 }
5423 else
5424 return 0;
5425 }
5426 }
5427 \f
5428 /* Help detect a giv that is calculated by several consecutive insns;
5429 for example,
5430 giv = biv * M
5431 giv = giv + A
5432 The caller has already identified the first insn P as having a giv as dest;
5433 we check that all other insns that set the same register follow
5434 immediately after P, that they alter nothing else,
5435 and that the result of the last is still a giv.
5436
5437 The value is 0 if the reg set in P is not really a giv.
5438 Otherwise, the value is the amount gained by eliminating
5439 all the consecutive insns that compute the value.
5440
5441 FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
5442 SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
5443
5444 The coefficients of the ultimate giv value are stored in
5445 *MULT_VAL and *ADD_VAL. */
5446
5447 static int
5448 consec_sets_giv (first_benefit, p, src_reg, dest_reg,
5449 add_val, mult_val)
5450 int first_benefit;
5451 rtx p;
5452 rtx src_reg;
5453 rtx dest_reg;
5454 rtx *add_val;
5455 rtx *mult_val;
5456 {
5457 int count;
5458 enum rtx_code code;
5459 int benefit;
5460 rtx temp;
5461 rtx set;
5462
5463 /* Indicate that this is a giv so that we can update the value produced in
5464 each insn of the multi-insn sequence.
5465
5466 This induction structure will be used only by the call to
5467 general_induction_var below, so we can allocate it on our stack.
5468 If this is a giv, our caller will replace the induct var entry with
5469 a new induction structure. */
5470 struct induction *v
5471 = (struct induction *) alloca (sizeof (struct induction));
5472 v->src_reg = src_reg;
5473 v->mult_val = *mult_val;
5474 v->add_val = *add_val;
5475 v->benefit = first_benefit;
5476 v->cant_derive = 0;
5477 v->derive_adjustment = 0;
5478
5479 reg_iv_type[REGNO (dest_reg)] = GENERAL_INDUCT;
5480 reg_iv_info[REGNO (dest_reg)] = v;
5481
5482 count = n_times_set[REGNO (dest_reg)] - 1;
5483
5484 while (count > 0)
5485 {
5486 p = NEXT_INSN (p);
5487 code = GET_CODE (p);
5488
5489 /* If libcall, skip to end of call sequence. */
5490 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
5491 p = XEXP (temp, 0);
5492
5493 if (code == INSN
5494 && (set = single_set (p))
5495 && GET_CODE (SET_DEST (set)) == REG
5496 && SET_DEST (set) == dest_reg
5497 && ((benefit = general_induction_var (SET_SRC (set), &src_reg,
5498 add_val, mult_val))
5499 /* Giv created by equivalent expression. */
5500 || ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
5501 && (benefit = general_induction_var (XEXP (temp, 0), &src_reg,
5502 add_val, mult_val))))
5503 && src_reg == v->src_reg)
5504 {
5505 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
5506 benefit += libcall_benefit (p);
5507
5508 count--;
5509 v->mult_val = *mult_val;
5510 v->add_val = *add_val;
5511 v->benefit = benefit;
5512 }
5513 else if (code != NOTE)
5514 {
5515 /* Allow insns that set something other than this giv to a
5516 constant. Such insns are needed on machines which cannot
5517 include long constants and should not disqualify a giv. */
5518 if (code == INSN
5519 && (set = single_set (p))
5520 && SET_DEST (set) != dest_reg
5521 && CONSTANT_P (SET_SRC (set)))
5522 continue;
5523
5524 reg_iv_type[REGNO (dest_reg)] = UNKNOWN_INDUCT;
5525 return 0;
5526 }
5527 }
5528
5529 return v->benefit;
5530 }
5531 \f
5532 /* Return an rtx, if any, that expresses giv G2 as a function of the register
5533 represented by G1. If no such expression can be found, or it is clear that
5534 it cannot possibly be a valid address, 0 is returned.
5535
5536 To perform the computation, we note that
5537 G1 = a * v + b and
5538 G2 = c * v + d
5539 where `v' is the biv.
5540
5541 So G2 = (c/a) * G1 + (d - b*c/a) */
5542
5543 #ifdef ADDRESS_COST
5544 static rtx
5545 express_from (g1, g2)
5546 struct induction *g1, *g2;
5547 {
5548 rtx mult, add;
5549
5550 /* The value that G1 will be multiplied by must be a constant integer. Also,
5551 the only chance we have of getting a valid address is if b*c/a (see above
5552 for notation) is also an integer. */
5553 if (GET_CODE (g1->mult_val) != CONST_INT
5554 || GET_CODE (g2->mult_val) != CONST_INT
5555 || GET_CODE (g1->add_val) != CONST_INT
5556 || g1->mult_val == const0_rtx
5557 || INTVAL (g2->mult_val) % INTVAL (g1->mult_val) != 0)
5558 return 0;
5559
5560 mult = GEN_INT (INTVAL (g2->mult_val) / INTVAL (g1->mult_val));
5561 add = plus_constant (g2->add_val, - INTVAL (g1->add_val) * INTVAL (mult));
5562
5563 /* Form simplified final result. */
5564 if (mult == const0_rtx)
5565 return add;
5566 else if (mult == const1_rtx)
5567 mult = g1->dest_reg;
5568 else
5569 mult = gen_rtx (MULT, g2->mode, g1->dest_reg, mult);
5570
5571 if (add == const0_rtx)
5572 return mult;
5573 else
5574 return gen_rtx (PLUS, g2->mode, mult, add);
5575 }
5576 #endif
5577 \f
5578 /* Return 1 if giv G2 can be combined with G1. This means that G2 can use
5579 (either directly or via an address expression) a register used to represent
5580 G1. Set g2->new_reg to a represtation of G1 (normally just
5581 g1->dest_reg). */
5582
5583 static int
5584 combine_givs_p (g1, g2)
5585 struct induction *g1, *g2;
5586 {
5587 rtx tem;
5588
5589 /* If these givs are identical, they can be combined. */
5590 if (rtx_equal_p (g1->mult_val, g2->mult_val)
5591 && rtx_equal_p (g1->add_val, g2->add_val))
5592 {
5593 g2->new_reg = g1->dest_reg;
5594 return 1;
5595 }
5596
5597 #ifdef ADDRESS_COST
5598 /* If G2 can be expressed as a function of G1 and that function is valid
5599 as an address and no more expensive than using a register for G2,
5600 the expression of G2 in terms of G1 can be used. */
5601 if (g2->giv_type == DEST_ADDR
5602 && (tem = express_from (g1, g2)) != 0
5603 && memory_address_p (g2->mem_mode, tem)
5604 && ADDRESS_COST (tem) <= ADDRESS_COST (*g2->location))
5605 {
5606 g2->new_reg = tem;
5607 return 1;
5608 }
5609 #endif
5610
5611 return 0;
5612 }
5613 \f
5614 #ifdef GIV_SORT_CRITERION
5615 /* Compare two givs and sort the most desirable one for combinations first.
5616 This is used only in one qsort call below. */
5617
5618 static int
5619 giv_sort (x, y)
5620 struct induction **x, **y;
5621 {
5622 GIV_SORT_CRITERION (*x, *y);
5623
5624 return 0;
5625 }
5626 #endif
5627
5628 /* Check all pairs of givs for iv_class BL and see if any can be combined with
5629 any other. If so, point SAME to the giv combined with and set NEW_REG to
5630 be an expression (in terms of the other giv's DEST_REG) equivalent to the
5631 giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
5632
5633 static void
5634 combine_givs (bl)
5635 struct iv_class *bl;
5636 {
5637 struct induction *g1, *g2, **giv_array, *temp_iv;
5638 int i, j, giv_count, pass;
5639
5640 /* Count givs, because bl->giv_count is incorrect here. */
5641 giv_count = 0;
5642 for (g1 = bl->giv; g1; g1 = g1->next_iv)
5643 giv_count++;
5644
5645 giv_array
5646 = (struct induction **) alloca (giv_count * sizeof (struct induction *));
5647 i = 0;
5648 for (g1 = bl->giv; g1; g1 = g1->next_iv)
5649 giv_array[i++] = g1;
5650
5651 #ifdef GIV_SORT_CRITERION
5652 /* Sort the givs if GIV_SORT_CRITERION is defined.
5653 This is usually defined for processors which lack
5654 negative register offsets so more givs may be combined. */
5655
5656 if (loop_dump_stream)
5657 fprintf (loop_dump_stream, "%d givs counted, sorting...\n", giv_count);
5658
5659 qsort (giv_array, giv_count, sizeof (struct induction *), giv_sort);
5660 #endif
5661
5662 for (i = 0; i < giv_count; i++)
5663 {
5664 g1 = giv_array[i];
5665 for (pass = 0; pass <= 1; pass++)
5666 for (j = 0; j < giv_count; j++)
5667 {
5668 g2 = giv_array[j];
5669 if (g1 != g2
5670 /* First try to combine with replaceable givs, then all givs. */
5671 && (g1->replaceable || pass == 1)
5672 /* If either has already been combined or is to be ignored, can't
5673 combine. */
5674 && ! g1->ignore && ! g2->ignore && ! g1->same && ! g2->same
5675 /* If something has been based on G2, G2 cannot itself be based
5676 on something else. */
5677 && ! g2->combined_with
5678 && combine_givs_p (g1, g2))
5679 {
5680 /* g2->new_reg set by `combine_givs_p' */
5681 g2->same = g1;
5682 g1->combined_with = 1;
5683
5684 /* If one of these givs is a DEST_REG that was only used
5685 once, by the other giv, this is actually a single use.
5686 The DEST_REG has the correct cost, while the other giv
5687 counts the REG use too often. */
5688 if (g2->giv_type == DEST_REG
5689 && n_times_used[REGNO (g2->dest_reg)] == 1
5690 && reg_mentioned_p (g2->dest_reg, PATTERN (g1->insn)))
5691 g1->benefit = g2->benefit;
5692 else if (g1->giv_type != DEST_REG
5693 || n_times_used[REGNO (g1->dest_reg)] != 1
5694 || ! reg_mentioned_p (g1->dest_reg,
5695 PATTERN (g2->insn)))
5696 {
5697 g1->benefit += g2->benefit;
5698 g1->times_used += g2->times_used;
5699 }
5700 /* ??? The new final_[bg]iv_value code does a much better job
5701 of finding replaceable giv's, and hence this code may no
5702 longer be necessary. */
5703 if (! g2->replaceable && REG_USERVAR_P (g2->dest_reg))
5704 g1->benefit -= copy_cost;
5705 g1->lifetime += g2->lifetime;
5706
5707 if (loop_dump_stream)
5708 fprintf (loop_dump_stream, "giv at %d combined with giv at %d\n",
5709 INSN_UID (g2->insn), INSN_UID (g1->insn));
5710 }
5711 }
5712 }
5713 }
5714 \f
5715 /* EMIT code before INSERT_BEFORE to set REG = B * M + A. */
5716
5717 void
5718 emit_iv_add_mult (b, m, a, reg, insert_before)
5719 rtx b; /* initial value of basic induction variable */
5720 rtx m; /* multiplicative constant */
5721 rtx a; /* additive constant */
5722 rtx reg; /* destination register */
5723 rtx insert_before;
5724 {
5725 rtx seq;
5726 rtx result;
5727
5728 /* Prevent unexpected sharing of these rtx. */
5729 a = copy_rtx (a);
5730 b = copy_rtx (b);
5731
5732 /* Increase the lifetime of any invariants moved further in code. */
5733 update_reg_last_use (a, insert_before);
5734 update_reg_last_use (b, insert_before);
5735 update_reg_last_use (m, insert_before);
5736
5737 start_sequence ();
5738 result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 0);
5739 if (reg != result)
5740 emit_move_insn (reg, result);
5741 seq = gen_sequence ();
5742 end_sequence ();
5743
5744 emit_insn_before (seq, insert_before);
5745 }
5746 \f
5747 /* Test whether A * B can be computed without
5748 an actual multiply insn. Value is 1 if so. */
5749
5750 static int
5751 product_cheap_p (a, b)
5752 rtx a;
5753 rtx b;
5754 {
5755 int i;
5756 rtx tmp;
5757 struct obstack *old_rtl_obstack = rtl_obstack;
5758 char *storage = (char *) obstack_alloc (&temp_obstack, 0);
5759 int win = 1;
5760
5761 /* If only one is constant, make it B. */
5762 if (GET_CODE (a) == CONST_INT)
5763 tmp = a, a = b, b = tmp;
5764
5765 /* If first constant, both constant, so don't need multiply. */
5766 if (GET_CODE (a) == CONST_INT)
5767 return 1;
5768
5769 /* If second not constant, neither is constant, so would need multiply. */
5770 if (GET_CODE (b) != CONST_INT)
5771 return 0;
5772
5773 /* One operand is constant, so might not need multiply insn. Generate the
5774 code for the multiply and see if a call or multiply, or long sequence
5775 of insns is generated. */
5776
5777 rtl_obstack = &temp_obstack;
5778 start_sequence ();
5779 expand_mult (GET_MODE (a), a, b, NULL_RTX, 0);
5780 tmp = gen_sequence ();
5781 end_sequence ();
5782
5783 if (GET_CODE (tmp) == SEQUENCE)
5784 {
5785 if (XVEC (tmp, 0) == 0)
5786 win = 1;
5787 else if (XVECLEN (tmp, 0) > 3)
5788 win = 0;
5789 else
5790 for (i = 0; i < XVECLEN (tmp, 0); i++)
5791 {
5792 rtx insn = XVECEXP (tmp, 0, i);
5793
5794 if (GET_CODE (insn) != INSN
5795 || (GET_CODE (PATTERN (insn)) == SET
5796 && GET_CODE (SET_SRC (PATTERN (insn))) == MULT)
5797 || (GET_CODE (PATTERN (insn)) == PARALLEL
5798 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET
5799 && GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == MULT))
5800 {
5801 win = 0;
5802 break;
5803 }
5804 }
5805 }
5806 else if (GET_CODE (tmp) == SET
5807 && GET_CODE (SET_SRC (tmp)) == MULT)
5808 win = 0;
5809 else if (GET_CODE (tmp) == PARALLEL
5810 && GET_CODE (XVECEXP (tmp, 0, 0)) == SET
5811 && GET_CODE (SET_SRC (XVECEXP (tmp, 0, 0))) == MULT)
5812 win = 0;
5813
5814 /* Free any storage we obtained in generating this multiply and restore rtl
5815 allocation to its normal obstack. */
5816 obstack_free (&temp_obstack, storage);
5817 rtl_obstack = old_rtl_obstack;
5818
5819 return win;
5820 }
5821 \f
5822 /* Check to see if loop can be terminated by a "decrement and branch until
5823 zero" instruction. If so, add a REG_NONNEG note to the branch insn if so.
5824 Also try reversing an increment loop to a decrement loop
5825 to see if the optimization can be performed.
5826 Value is nonzero if optimization was performed. */
5827
5828 /* This is useful even if the architecture doesn't have such an insn,
5829 because it might change a loops which increments from 0 to n to a loop
5830 which decrements from n to 0. A loop that decrements to zero is usually
5831 faster than one that increments from zero. */
5832
5833 /* ??? This could be rewritten to use some of the loop unrolling procedures,
5834 such as approx_final_value, biv_total_increment, loop_iterations, and
5835 final_[bg]iv_value. */
5836
5837 static int
5838 check_dbra_loop (loop_end, insn_count, loop_start)
5839 rtx loop_end;
5840 int insn_count;
5841 rtx loop_start;
5842 {
5843 struct iv_class *bl;
5844 rtx reg;
5845 rtx jump_label;
5846 rtx final_value;
5847 rtx start_value;
5848 rtx new_add_val;
5849 rtx comparison;
5850 rtx before_comparison;
5851 rtx p;
5852
5853 /* If last insn is a conditional branch, and the insn before tests a
5854 register value, try to optimize it. Otherwise, we can't do anything. */
5855
5856 comparison = get_condition_for_loop (PREV_INSN (loop_end));
5857 if (comparison == 0)
5858 return 0;
5859
5860 /* Check all of the bivs to see if the compare uses one of them.
5861 Skip biv's set more than once because we can't guarantee that
5862 it will be zero on the last iteration. Also skip if the biv is
5863 used between its update and the test insn. */
5864
5865 for (bl = loop_iv_list; bl; bl = bl->next)
5866 {
5867 if (bl->biv_count == 1
5868 && bl->biv->dest_reg == XEXP (comparison, 0)
5869 && ! reg_used_between_p (regno_reg_rtx[bl->regno], bl->biv->insn,
5870 PREV_INSN (PREV_INSN (loop_end))))
5871 break;
5872 }
5873
5874 if (! bl)
5875 return 0;
5876
5877 /* Look for the case where the basic induction variable is always
5878 nonnegative, and equals zero on the last iteration.
5879 In this case, add a reg_note REG_NONNEG, which allows the
5880 m68k DBRA instruction to be used. */
5881
5882 if (((GET_CODE (comparison) == GT
5883 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
5884 && INTVAL (XEXP (comparison, 1)) == -1)
5885 || (GET_CODE (comparison) == NE && XEXP (comparison, 1) == const0_rtx))
5886 && GET_CODE (bl->biv->add_val) == CONST_INT
5887 && INTVAL (bl->biv->add_val) < 0)
5888 {
5889 /* Initial value must be greater than 0,
5890 init_val % -dec_value == 0 to ensure that it equals zero on
5891 the last iteration */
5892
5893 if (GET_CODE (bl->initial_value) == CONST_INT
5894 && INTVAL (bl->initial_value) > 0
5895 && (INTVAL (bl->initial_value) %
5896 (-INTVAL (bl->biv->add_val))) == 0)
5897 {
5898 /* register always nonnegative, add REG_NOTE to branch */
5899 REG_NOTES (PREV_INSN (loop_end))
5900 = gen_rtx (EXPR_LIST, REG_NONNEG, NULL_RTX,
5901 REG_NOTES (PREV_INSN (loop_end)));
5902 bl->nonneg = 1;
5903
5904 return 1;
5905 }
5906
5907 /* If the decrement is 1 and the value was tested as >= 0 before
5908 the loop, then we can safely optimize. */
5909 for (p = loop_start; p; p = PREV_INSN (p))
5910 {
5911 if (GET_CODE (p) == CODE_LABEL)
5912 break;
5913 if (GET_CODE (p) != JUMP_INSN)
5914 continue;
5915
5916 before_comparison = get_condition_for_loop (p);
5917 if (before_comparison
5918 && XEXP (before_comparison, 0) == bl->biv->dest_reg
5919 && GET_CODE (before_comparison) == LT
5920 && XEXP (before_comparison, 1) == const0_rtx
5921 && ! reg_set_between_p (bl->biv->dest_reg, p, loop_start)
5922 && INTVAL (bl->biv->add_val) == -1)
5923 {
5924 REG_NOTES (PREV_INSN (loop_end))
5925 = gen_rtx (EXPR_LIST, REG_NONNEG, NULL_RTX,
5926 REG_NOTES (PREV_INSN (loop_end)));
5927 bl->nonneg = 1;
5928
5929 return 1;
5930 }
5931 }
5932 }
5933 else if (num_mem_sets <= 1)
5934 {
5935 /* Try to change inc to dec, so can apply above optimization. */
5936 /* Can do this if:
5937 all registers modified are induction variables or invariant,
5938 all memory references have non-overlapping addresses
5939 (obviously true if only one write)
5940 allow 2 insns for the compare/jump at the end of the loop. */
5941 /* Also, we must avoid any instructions which use both the reversed
5942 biv and another biv. Such instructions will fail if the loop is
5943 reversed. We meet this condition by requiring that either
5944 no_use_except_counting is true, or else that there is only
5945 one biv. */
5946 int num_nonfixed_reads = 0;
5947 /* 1 if the iteration var is used only to count iterations. */
5948 int no_use_except_counting = 0;
5949 /* 1 if the loop has no memory store, or it has a single memory store
5950 which is reversible. */
5951 int reversible_mem_store = 1;
5952
5953 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
5954 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
5955 num_nonfixed_reads += count_nonfixed_reads (PATTERN (p));
5956
5957 if (bl->giv_count == 0
5958 && ! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
5959 {
5960 rtx bivreg = regno_reg_rtx[bl->regno];
5961
5962 /* If there are no givs for this biv, and the only exit is the
5963 fall through at the end of the the loop, then
5964 see if perhaps there are no uses except to count. */
5965 no_use_except_counting = 1;
5966 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
5967 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
5968 {
5969 rtx set = single_set (p);
5970
5971 if (set && GET_CODE (SET_DEST (set)) == REG
5972 && REGNO (SET_DEST (set)) == bl->regno)
5973 /* An insn that sets the biv is okay. */
5974 ;
5975 else if (p == prev_nonnote_insn (prev_nonnote_insn (loop_end))
5976 || p == prev_nonnote_insn (loop_end))
5977 /* Don't bother about the end test. */
5978 ;
5979 else if (reg_mentioned_p (bivreg, PATTERN (p)))
5980 /* Any other use of the biv is no good. */
5981 {
5982 no_use_except_counting = 0;
5983 break;
5984 }
5985 }
5986 }
5987
5988 /* If the loop has a single store, and the destination address is
5989 invariant, then we can't reverse the loop, because this address
5990 might then have the wrong value at loop exit.
5991 This would work if the source was invariant also, however, in that
5992 case, the insn should have been moved out of the loop. */
5993
5994 if (num_mem_sets == 1)
5995 reversible_mem_store
5996 = (! unknown_address_altered
5997 && ! invariant_p (XEXP (loop_store_mems[0], 0)));
5998
5999 /* This code only acts for innermost loops. Also it simplifies
6000 the memory address check by only reversing loops with
6001 zero or one memory access.
6002 Two memory accesses could involve parts of the same array,
6003 and that can't be reversed. */
6004
6005 if (num_nonfixed_reads <= 1
6006 && !loop_has_call
6007 && !loop_has_volatile
6008 && reversible_mem_store
6009 && (no_use_except_counting
6010 || ((bl->giv_count + bl->biv_count + num_mem_sets
6011 + num_movables + 2 == insn_count)
6012 && (bl == loop_iv_list && bl->next == 0))))
6013 {
6014 rtx tem;
6015
6016 /* Loop can be reversed. */
6017 if (loop_dump_stream)
6018 fprintf (loop_dump_stream, "Can reverse loop\n");
6019
6020 /* Now check other conditions:
6021 initial_value must be zero,
6022 final_value % add_val == 0, so that when reversed, the
6023 biv will be zero on the last iteration.
6024
6025 This test can probably be improved since +/- 1 in the constant
6026 can be obtained by changing LT to LE and vice versa; this is
6027 confusing. */
6028
6029 if (comparison && bl->initial_value == const0_rtx
6030 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
6031 /* LE gets turned into LT */
6032 && GET_CODE (comparison) == LT
6033 && (INTVAL (XEXP (comparison, 1))
6034 % INTVAL (bl->biv->add_val)) == 0)
6035 {
6036 /* Register will always be nonnegative, with value
6037 0 on last iteration if loop reversed */
6038
6039 /* Save some info needed to produce the new insns. */
6040 reg = bl->biv->dest_reg;
6041 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 1);
6042 if (jump_label == pc_rtx)
6043 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 2);
6044 new_add_val = GEN_INT (- INTVAL (bl->biv->add_val));
6045
6046 final_value = XEXP (comparison, 1);
6047 start_value = GEN_INT (INTVAL (XEXP (comparison, 1))
6048 - INTVAL (bl->biv->add_val));
6049
6050 /* Initialize biv to start_value before loop start.
6051 The old initializing insn will be deleted as a
6052 dead store by flow.c. */
6053 emit_insn_before (gen_move_insn (reg, start_value), loop_start);
6054
6055 /* Add insn to decrement register, and delete insn
6056 that incremented the register. */
6057 p = emit_insn_before (gen_add2_insn (reg, new_add_val),
6058 bl->biv->insn);
6059 delete_insn (bl->biv->insn);
6060
6061 /* Update biv info to reflect its new status. */
6062 bl->biv->insn = p;
6063 bl->initial_value = start_value;
6064 bl->biv->add_val = new_add_val;
6065
6066 /* Inc LABEL_NUSES so that delete_insn will
6067 not delete the label. */
6068 LABEL_NUSES (XEXP (jump_label, 0)) ++;
6069
6070 /* Emit an insn after the end of the loop to set the biv's
6071 proper exit value if it is used anywhere outside the loop. */
6072 if ((regno_last_uid[bl->regno]
6073 != INSN_UID (PREV_INSN (PREV_INSN (loop_end))))
6074 || ! bl->init_insn
6075 || regno_first_uid[bl->regno] != INSN_UID (bl->init_insn))
6076 emit_insn_after (gen_move_insn (reg, final_value),
6077 loop_end);
6078
6079 /* Delete compare/branch at end of loop. */
6080 delete_insn (PREV_INSN (loop_end));
6081 delete_insn (PREV_INSN (loop_end));
6082
6083 /* Add new compare/branch insn at end of loop. */
6084 start_sequence ();
6085 emit_cmp_insn (reg, const0_rtx, GE, NULL_RTX,
6086 GET_MODE (reg), 0, 0);
6087 emit_jump_insn (gen_bge (XEXP (jump_label, 0)));
6088 tem = gen_sequence ();
6089 end_sequence ();
6090 emit_jump_insn_before (tem, loop_end);
6091
6092 for (tem = PREV_INSN (loop_end);
6093 tem && GET_CODE (tem) != JUMP_INSN; tem = PREV_INSN (tem))
6094 ;
6095 if (tem)
6096 {
6097 JUMP_LABEL (tem) = XEXP (jump_label, 0);
6098
6099 /* Increment of LABEL_NUSES done above. */
6100 /* Register is now always nonnegative,
6101 so add REG_NONNEG note to the branch. */
6102 REG_NOTES (tem) = gen_rtx (EXPR_LIST, REG_NONNEG, NULL_RTX,
6103 REG_NOTES (tem));
6104 }
6105
6106 bl->nonneg = 1;
6107
6108 /* Mark that this biv has been reversed. Each giv which depends
6109 on this biv, and which is also live past the end of the loop
6110 will have to be fixed up. */
6111
6112 bl->reversed = 1;
6113
6114 if (loop_dump_stream)
6115 fprintf (loop_dump_stream,
6116 "Reversed loop and added reg_nonneg\n");
6117
6118 return 1;
6119 }
6120 }
6121 }
6122
6123 return 0;
6124 }
6125 \f
6126 /* Verify whether the biv BL appears to be eliminable,
6127 based on the insns in the loop that refer to it.
6128 LOOP_START is the first insn of the loop, and END is the end insn.
6129
6130 If ELIMINATE_P is non-zero, actually do the elimination.
6131
6132 THRESHOLD and INSN_COUNT are from loop_optimize and are used to
6133 determine whether invariant insns should be placed inside or at the
6134 start of the loop. */
6135
6136 static int
6137 maybe_eliminate_biv (bl, loop_start, end, eliminate_p, threshold, insn_count)
6138 struct iv_class *bl;
6139 rtx loop_start;
6140 rtx end;
6141 int eliminate_p;
6142 int threshold, insn_count;
6143 {
6144 rtx reg = bl->biv->dest_reg;
6145 rtx p;
6146
6147 /* Scan all insns in the loop, stopping if we find one that uses the
6148 biv in a way that we cannot eliminate. */
6149
6150 for (p = loop_start; p != end; p = NEXT_INSN (p))
6151 {
6152 enum rtx_code code = GET_CODE (p);
6153 rtx where = threshold >= insn_count ? loop_start : p;
6154
6155 if ((code == INSN || code == JUMP_INSN || code == CALL_INSN)
6156 && reg_mentioned_p (reg, PATTERN (p))
6157 && ! maybe_eliminate_biv_1 (PATTERN (p), p, bl, eliminate_p, where))
6158 {
6159 if (loop_dump_stream)
6160 fprintf (loop_dump_stream,
6161 "Cannot eliminate biv %d: biv used in insn %d.\n",
6162 bl->regno, INSN_UID (p));
6163 break;
6164 }
6165 }
6166
6167 if (p == end)
6168 {
6169 if (loop_dump_stream)
6170 fprintf (loop_dump_stream, "biv %d %s eliminated.\n",
6171 bl->regno, eliminate_p ? "was" : "can be");
6172 return 1;
6173 }
6174
6175 return 0;
6176 }
6177 \f
6178 /* If BL appears in X (part of the pattern of INSN), see if we can
6179 eliminate its use. If so, return 1. If not, return 0.
6180
6181 If BIV does not appear in X, return 1.
6182
6183 If ELIMINATE_P is non-zero, actually do the elimination. WHERE indicates
6184 where extra insns should be added. Depending on how many items have been
6185 moved out of the loop, it will either be before INSN or at the start of
6186 the loop. */
6187
6188 static int
6189 maybe_eliminate_biv_1 (x, insn, bl, eliminate_p, where)
6190 rtx x, insn;
6191 struct iv_class *bl;
6192 int eliminate_p;
6193 rtx where;
6194 {
6195 enum rtx_code code = GET_CODE (x);
6196 rtx reg = bl->biv->dest_reg;
6197 enum machine_mode mode = GET_MODE (reg);
6198 struct induction *v;
6199 rtx arg, new, tem;
6200 int arg_operand;
6201 char *fmt;
6202 int i, j;
6203
6204 switch (code)
6205 {
6206 case REG:
6207 /* If we haven't already been able to do something with this BIV,
6208 we can't eliminate it. */
6209 if (x == reg)
6210 return 0;
6211 return 1;
6212
6213 case SET:
6214 /* If this sets the BIV, it is not a problem. */
6215 if (SET_DEST (x) == reg)
6216 return 1;
6217
6218 /* If this is an insn that defines a giv, it is also ok because
6219 it will go away when the giv is reduced. */
6220 for (v = bl->giv; v; v = v->next_iv)
6221 if (v->giv_type == DEST_REG && SET_DEST (x) == v->dest_reg)
6222 return 1;
6223
6224 #ifdef HAVE_cc0
6225 if (SET_DEST (x) == cc0_rtx && SET_SRC (x) == reg)
6226 {
6227 /* Can replace with any giv that was reduced and
6228 that has (MULT_VAL != 0) and (ADD_VAL == 0).
6229 Require a constant for MULT_VAL, so we know it's nonzero.
6230 ??? We disable this optimization to avoid potential
6231 overflows. */
6232
6233 for (v = bl->giv; v; v = v->next_iv)
6234 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
6235 && v->add_val == const0_rtx
6236 && ! v->ignore && ! v->maybe_dead && v->always_computable
6237 && v->mode == mode
6238 && 0)
6239 {
6240 /* If the giv V had the auto-inc address optimization applied
6241 to it, and INSN occurs between the giv insn and the biv
6242 insn, then we must adjust the value used here.
6243 This is rare, so we don't bother to do so. */
6244 if (v->auto_inc_opt
6245 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6246 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6247 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6248 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6249 continue;
6250
6251 if (! eliminate_p)
6252 return 1;
6253
6254 /* If the giv has the opposite direction of change,
6255 then reverse the comparison. */
6256 if (INTVAL (v->mult_val) < 0)
6257 new = gen_rtx (COMPARE, GET_MODE (v->new_reg),
6258 const0_rtx, v->new_reg);
6259 else
6260 new = v->new_reg;
6261
6262 /* We can probably test that giv's reduced reg. */
6263 if (validate_change (insn, &SET_SRC (x), new, 0))
6264 return 1;
6265 }
6266
6267 /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
6268 replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
6269 Require a constant for MULT_VAL, so we know it's nonzero.
6270 ??? Do this only if ADD_VAL is a pointer to avoid a potential
6271 overflow problem. */
6272
6273 for (v = bl->giv; v; v = v->next_iv)
6274 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
6275 && ! v->ignore && ! v->maybe_dead && v->always_computable
6276 && v->mode == mode
6277 && (GET_CODE (v->add_val) == SYMBOL_REF
6278 || GET_CODE (v->add_val) == LABEL_REF
6279 || GET_CODE (v->add_val) == CONST
6280 || (GET_CODE (v->add_val) == REG
6281 && REGNO_POINTER_FLAG (REGNO (v->add_val)))))
6282 {
6283 /* If the giv V had the auto-inc address optimization applied
6284 to it, and INSN occurs between the giv insn and the biv
6285 insn, then we must adjust the value used here.
6286 This is rare, so we don't bother to do so. */
6287 if (v->auto_inc_opt
6288 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6289 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6290 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6291 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6292 continue;
6293
6294 if (! eliminate_p)
6295 return 1;
6296
6297 /* If the giv has the opposite direction of change,
6298 then reverse the comparison. */
6299 if (INTVAL (v->mult_val) < 0)
6300 new = gen_rtx (COMPARE, VOIDmode, copy_rtx (v->add_val),
6301 v->new_reg);
6302 else
6303 new = gen_rtx (COMPARE, VOIDmode, v->new_reg,
6304 copy_rtx (v->add_val));
6305
6306 /* Replace biv with the giv's reduced register. */
6307 update_reg_last_use (v->add_val, insn);
6308 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
6309 return 1;
6310
6311 /* Insn doesn't support that constant or invariant. Copy it
6312 into a register (it will be a loop invariant.) */
6313 tem = gen_reg_rtx (GET_MODE (v->new_reg));
6314
6315 emit_insn_before (gen_move_insn (tem, copy_rtx (v->add_val)),
6316 where);
6317
6318 if (validate_change (insn, &SET_SRC (PATTERN (insn)),
6319 gen_rtx (COMPARE, VOIDmode,
6320 v->new_reg, tem), 0))
6321 return 1;
6322 }
6323 }
6324 #endif
6325 break;
6326
6327 case COMPARE:
6328 case EQ: case NE:
6329 case GT: case GE: case GTU: case GEU:
6330 case LT: case LE: case LTU: case LEU:
6331 /* See if either argument is the biv. */
6332 if (XEXP (x, 0) == reg)
6333 arg = XEXP (x, 1), arg_operand = 1;
6334 else if (XEXP (x, 1) == reg)
6335 arg = XEXP (x, 0), arg_operand = 0;
6336 else
6337 break;
6338
6339 if (CONSTANT_P (arg))
6340 {
6341 /* First try to replace with any giv that has constant positive
6342 mult_val and constant add_val. We might be able to support
6343 negative mult_val, but it seems complex to do it in general. */
6344
6345 for (v = bl->giv; v; v = v->next_iv)
6346 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
6347 && (GET_CODE (v->add_val) == SYMBOL_REF
6348 || GET_CODE (v->add_val) == LABEL_REF
6349 || GET_CODE (v->add_val) == CONST
6350 || (GET_CODE (v->add_val) == REG
6351 && REGNO_POINTER_FLAG (REGNO (v->add_val))))
6352 && ! v->ignore && ! v->maybe_dead && v->always_computable
6353 && v->mode == mode)
6354 {
6355 /* If the giv V had the auto-inc address optimization applied
6356 to it, and INSN occurs between the giv insn and the biv
6357 insn, then we must adjust the value used here.
6358 This is rare, so we don't bother to do so. */
6359 if (v->auto_inc_opt
6360 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6361 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6362 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6363 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6364 continue;
6365
6366 if (! eliminate_p)
6367 return 1;
6368
6369 /* Replace biv with the giv's reduced reg. */
6370 XEXP (x, 1-arg_operand) = v->new_reg;
6371
6372 /* If all constants are actually constant integers and
6373 the derived constant can be directly placed in the COMPARE,
6374 do so. */
6375 if (GET_CODE (arg) == CONST_INT
6376 && GET_CODE (v->mult_val) == CONST_INT
6377 && GET_CODE (v->add_val) == CONST_INT
6378 && validate_change (insn, &XEXP (x, arg_operand),
6379 GEN_INT (INTVAL (arg)
6380 * INTVAL (v->mult_val)
6381 + INTVAL (v->add_val)), 0))
6382 return 1;
6383
6384 /* Otherwise, load it into a register. */
6385 tem = gen_reg_rtx (mode);
6386 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
6387 if (validate_change (insn, &XEXP (x, arg_operand), tem, 0))
6388 return 1;
6389
6390 /* If that failed, put back the change we made above. */
6391 XEXP (x, 1-arg_operand) = reg;
6392 }
6393
6394 /* Look for giv with positive constant mult_val and nonconst add_val.
6395 Insert insns to calculate new compare value.
6396 ??? Turn this off due to possible overflow. */
6397
6398 for (v = bl->giv; v; v = v->next_iv)
6399 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
6400 && ! v->ignore && ! v->maybe_dead && v->always_computable
6401 && v->mode == mode
6402 && 0)
6403 {
6404 rtx tem;
6405
6406 /* If the giv V had the auto-inc address optimization applied
6407 to it, and INSN occurs between the giv insn and the biv
6408 insn, then we must adjust the value used here.
6409 This is rare, so we don't bother to do so. */
6410 if (v->auto_inc_opt
6411 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6412 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6413 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6414 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6415 continue;
6416
6417 if (! eliminate_p)
6418 return 1;
6419
6420 tem = gen_reg_rtx (mode);
6421
6422 /* Replace biv with giv's reduced register. */
6423 validate_change (insn, &XEXP (x, 1 - arg_operand),
6424 v->new_reg, 1);
6425
6426 /* Compute value to compare against. */
6427 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
6428 /* Use it in this insn. */
6429 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
6430 if (apply_change_group ())
6431 return 1;
6432 }
6433 }
6434 else if (GET_CODE (arg) == REG || GET_CODE (arg) == MEM)
6435 {
6436 if (invariant_p (arg) == 1)
6437 {
6438 /* Look for giv with constant positive mult_val and nonconst
6439 add_val. Insert insns to compute new compare value.
6440 ??? Turn this off due to possible overflow. */
6441
6442 for (v = bl->giv; v; v = v->next_iv)
6443 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
6444 && ! v->ignore && ! v->maybe_dead && v->always_computable
6445 && v->mode == mode
6446 && 0)
6447 {
6448 rtx tem;
6449
6450 /* If the giv V had the auto-inc address optimization applied
6451 to it, and INSN occurs between the giv insn and the biv
6452 insn, then we must adjust the value used here.
6453 This is rare, so we don't bother to do so. */
6454 if (v->auto_inc_opt
6455 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6456 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6457 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6458 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6459 continue;
6460
6461 if (! eliminate_p)
6462 return 1;
6463
6464 tem = gen_reg_rtx (mode);
6465
6466 /* Replace biv with giv's reduced register. */
6467 validate_change (insn, &XEXP (x, 1 - arg_operand),
6468 v->new_reg, 1);
6469
6470 /* Compute value to compare against. */
6471 emit_iv_add_mult (arg, v->mult_val, v->add_val,
6472 tem, where);
6473 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
6474 if (apply_change_group ())
6475 return 1;
6476 }
6477 }
6478
6479 /* This code has problems. Basically, you can't know when
6480 seeing if we will eliminate BL, whether a particular giv
6481 of ARG will be reduced. If it isn't going to be reduced,
6482 we can't eliminate BL. We can try forcing it to be reduced,
6483 but that can generate poor code.
6484
6485 The problem is that the benefit of reducing TV, below should
6486 be increased if BL can actually be eliminated, but this means
6487 we might have to do a topological sort of the order in which
6488 we try to process biv. It doesn't seem worthwhile to do
6489 this sort of thing now. */
6490
6491 #if 0
6492 /* Otherwise the reg compared with had better be a biv. */
6493 if (GET_CODE (arg) != REG
6494 || reg_iv_type[REGNO (arg)] != BASIC_INDUCT)
6495 return 0;
6496
6497 /* Look for a pair of givs, one for each biv,
6498 with identical coefficients. */
6499 for (v = bl->giv; v; v = v->next_iv)
6500 {
6501 struct induction *tv;
6502
6503 if (v->ignore || v->maybe_dead || v->mode != mode)
6504 continue;
6505
6506 for (tv = reg_biv_class[REGNO (arg)]->giv; tv; tv = tv->next_iv)
6507 if (! tv->ignore && ! tv->maybe_dead
6508 && rtx_equal_p (tv->mult_val, v->mult_val)
6509 && rtx_equal_p (tv->add_val, v->add_val)
6510 && tv->mode == mode)
6511 {
6512 /* If the giv V had the auto-inc address optimization applied
6513 to it, and INSN occurs between the giv insn and the biv
6514 insn, then we must adjust the value used here.
6515 This is rare, so we don't bother to do so. */
6516 if (v->auto_inc_opt
6517 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6518 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6519 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6520 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6521 continue;
6522
6523 if (! eliminate_p)
6524 return 1;
6525
6526 /* Replace biv with its giv's reduced reg. */
6527 XEXP (x, 1-arg_operand) = v->new_reg;
6528 /* Replace other operand with the other giv's
6529 reduced reg. */
6530 XEXP (x, arg_operand) = tv->new_reg;
6531 return 1;
6532 }
6533 }
6534 #endif
6535 }
6536
6537 /* If we get here, the biv can't be eliminated. */
6538 return 0;
6539
6540 case MEM:
6541 /* If this address is a DEST_ADDR giv, it doesn't matter if the
6542 biv is used in it, since it will be replaced. */
6543 for (v = bl->giv; v; v = v->next_iv)
6544 if (v->giv_type == DEST_ADDR && v->location == &XEXP (x, 0))
6545 return 1;
6546 break;
6547 }
6548
6549 /* See if any subexpression fails elimination. */
6550 fmt = GET_RTX_FORMAT (code);
6551 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6552 {
6553 switch (fmt[i])
6554 {
6555 case 'e':
6556 if (! maybe_eliminate_biv_1 (XEXP (x, i), insn, bl,
6557 eliminate_p, where))
6558 return 0;
6559 break;
6560
6561 case 'E':
6562 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6563 if (! maybe_eliminate_biv_1 (XVECEXP (x, i, j), insn, bl,
6564 eliminate_p, where))
6565 return 0;
6566 break;
6567 }
6568 }
6569
6570 return 1;
6571 }
6572 \f
6573 /* Return nonzero if the last use of REG
6574 is in an insn following INSN in the same basic block. */
6575
6576 static int
6577 last_use_this_basic_block (reg, insn)
6578 rtx reg;
6579 rtx insn;
6580 {
6581 rtx n;
6582 for (n = insn;
6583 n && GET_CODE (n) != CODE_LABEL && GET_CODE (n) != JUMP_INSN;
6584 n = NEXT_INSN (n))
6585 {
6586 if (regno_last_uid[REGNO (reg)] == INSN_UID (n))
6587 return 1;
6588 }
6589 return 0;
6590 }
6591 \f
6592 /* Called via `note_stores' to record the initial value of a biv. Here we
6593 just record the location of the set and process it later. */
6594
6595 static void
6596 record_initial (dest, set)
6597 rtx dest;
6598 rtx set;
6599 {
6600 struct iv_class *bl;
6601
6602 if (GET_CODE (dest) != REG
6603 || REGNO (dest) >= max_reg_before_loop
6604 || reg_iv_type[REGNO (dest)] != BASIC_INDUCT)
6605 return;
6606
6607 bl = reg_biv_class[REGNO (dest)];
6608
6609 /* If this is the first set found, record it. */
6610 if (bl->init_insn == 0)
6611 {
6612 bl->init_insn = note_insn;
6613 bl->init_set = set;
6614 }
6615 }
6616 \f
6617 /* If any of the registers in X are "old" and currently have a last use earlier
6618 than INSN, update them to have a last use of INSN. Their actual last use
6619 will be the previous insn but it will not have a valid uid_luid so we can't
6620 use it. */
6621
6622 static void
6623 update_reg_last_use (x, insn)
6624 rtx x;
6625 rtx insn;
6626 {
6627 /* Check for the case where INSN does not have a valid luid. In this case,
6628 there is no need to modify the regno_last_uid, as this can only happen
6629 when code is inserted after the loop_end to set a pseudo's final value,
6630 and hence this insn will never be the last use of x. */
6631 if (GET_CODE (x) == REG && REGNO (x) < max_reg_before_loop
6632 && INSN_UID (insn) < max_uid_for_loop
6633 && uid_luid[regno_last_uid[REGNO (x)]] < uid_luid[INSN_UID (insn)])
6634 regno_last_uid[REGNO (x)] = INSN_UID (insn);
6635 else
6636 {
6637 register int i, j;
6638 register char *fmt = GET_RTX_FORMAT (GET_CODE (x));
6639 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
6640 {
6641 if (fmt[i] == 'e')
6642 update_reg_last_use (XEXP (x, i), insn);
6643 else if (fmt[i] == 'E')
6644 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6645 update_reg_last_use (XVECEXP (x, i, j), insn);
6646 }
6647 }
6648 }
6649 \f
6650 /* Given a jump insn JUMP, return the condition that will cause it to branch
6651 to its JUMP_LABEL. If the condition cannot be understood, or is an
6652 inequality floating-point comparison which needs to be reversed, 0 will
6653 be returned.
6654
6655 If EARLIEST is non-zero, it is a pointer to a place where the earliest
6656 insn used in locating the condition was found. If a replacement test
6657 of the condition is desired, it should be placed in front of that
6658 insn and we will be sure that the inputs are still valid.
6659
6660 The condition will be returned in a canonical form to simplify testing by
6661 callers. Specifically:
6662
6663 (1) The code will always be a comparison operation (EQ, NE, GT, etc.).
6664 (2) Both operands will be machine operands; (cc0) will have been replaced.
6665 (3) If an operand is a constant, it will be the second operand.
6666 (4) (LE x const) will be replaced with (LT x <const+1>) and similarly
6667 for GE, GEU, and LEU. */
6668
6669 rtx
6670 get_condition (jump, earliest)
6671 rtx jump;
6672 rtx *earliest;
6673 {
6674 enum rtx_code code;
6675 rtx prev = jump;
6676 rtx set;
6677 rtx tem;
6678 rtx op0, op1;
6679 int reverse_code = 0;
6680 int did_reverse_condition = 0;
6681
6682 /* If this is not a standard conditional jump, we can't parse it. */
6683 if (GET_CODE (jump) != JUMP_INSN
6684 || ! condjump_p (jump) || simplejump_p (jump))
6685 return 0;
6686
6687 code = GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 0));
6688 op0 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 0);
6689 op1 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 1);
6690
6691 if (earliest)
6692 *earliest = jump;
6693
6694 /* If this branches to JUMP_LABEL when the condition is false, reverse
6695 the condition. */
6696 if (GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 2)) == LABEL_REF
6697 && XEXP (XEXP (SET_SRC (PATTERN (jump)), 2), 0) == JUMP_LABEL (jump))
6698 code = reverse_condition (code), did_reverse_condition ^= 1;
6699
6700 /* If we are comparing a register with zero, see if the register is set
6701 in the previous insn to a COMPARE or a comparison operation. Perform
6702 the same tests as a function of STORE_FLAG_VALUE as find_comparison_args
6703 in cse.c */
6704
6705 while (GET_RTX_CLASS (code) == '<' && op1 == CONST0_RTX (GET_MODE (op0)))
6706 {
6707 /* Set non-zero when we find something of interest. */
6708 rtx x = 0;
6709
6710 #ifdef HAVE_cc0
6711 /* If comparison with cc0, import actual comparison from compare
6712 insn. */
6713 if (op0 == cc0_rtx)
6714 {
6715 if ((prev = prev_nonnote_insn (prev)) == 0
6716 || GET_CODE (prev) != INSN
6717 || (set = single_set (prev)) == 0
6718 || SET_DEST (set) != cc0_rtx)
6719 return 0;
6720
6721 op0 = SET_SRC (set);
6722 op1 = CONST0_RTX (GET_MODE (op0));
6723 if (earliest)
6724 *earliest = prev;
6725 }
6726 #endif
6727
6728 /* If this is a COMPARE, pick up the two things being compared. */
6729 if (GET_CODE (op0) == COMPARE)
6730 {
6731 op1 = XEXP (op0, 1);
6732 op0 = XEXP (op0, 0);
6733 continue;
6734 }
6735 else if (GET_CODE (op0) != REG)
6736 break;
6737
6738 /* Go back to the previous insn. Stop if it is not an INSN. We also
6739 stop if it isn't a single set or if it has a REG_INC note because
6740 we don't want to bother dealing with it. */
6741
6742 if ((prev = prev_nonnote_insn (prev)) == 0
6743 || GET_CODE (prev) != INSN
6744 || FIND_REG_INC_NOTE (prev, 0)
6745 || (set = single_set (prev)) == 0)
6746 break;
6747
6748 /* If this is setting OP0, get what it sets it to if it looks
6749 relevant. */
6750 if (rtx_equal_p (SET_DEST (set), op0))
6751 {
6752 enum machine_mode inner_mode = GET_MODE (SET_SRC (set));
6753
6754 if ((GET_CODE (SET_SRC (set)) == COMPARE
6755 || (((code == NE
6756 || (code == LT
6757 && GET_MODE_CLASS (inner_mode) == MODE_INT
6758 && (GET_MODE_BITSIZE (inner_mode)
6759 <= HOST_BITS_PER_WIDE_INT)
6760 && (STORE_FLAG_VALUE
6761 & ((HOST_WIDE_INT) 1
6762 << (GET_MODE_BITSIZE (inner_mode) - 1))))
6763 #ifdef FLOAT_STORE_FLAG_VALUE
6764 || (code == LT
6765 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
6766 && FLOAT_STORE_FLAG_VALUE < 0)
6767 #endif
6768 ))
6769 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<')))
6770 x = SET_SRC (set);
6771 else if (((code == EQ
6772 || (code == GE
6773 && (GET_MODE_BITSIZE (inner_mode)
6774 <= HOST_BITS_PER_WIDE_INT)
6775 && GET_MODE_CLASS (inner_mode) == MODE_INT
6776 && (STORE_FLAG_VALUE
6777 & ((HOST_WIDE_INT) 1
6778 << (GET_MODE_BITSIZE (inner_mode) - 1))))
6779 #ifdef FLOAT_STORE_FLAG_VALUE
6780 || (code == GE
6781 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
6782 && FLOAT_STORE_FLAG_VALUE < 0)
6783 #endif
6784 ))
6785 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<')
6786 {
6787 /* We might have reversed a LT to get a GE here. But this wasn't
6788 actually the comparison of data, so we don't flag that we
6789 have had to reverse the condition. */
6790 did_reverse_condition ^= 1;
6791 reverse_code = 1;
6792 x = SET_SRC (set);
6793 }
6794 else
6795 break;
6796 }
6797
6798 else if (reg_set_p (op0, prev))
6799 /* If this sets OP0, but not directly, we have to give up. */
6800 break;
6801
6802 if (x)
6803 {
6804 if (GET_RTX_CLASS (GET_CODE (x)) == '<')
6805 code = GET_CODE (x);
6806 if (reverse_code)
6807 {
6808 code = reverse_condition (code);
6809 did_reverse_condition ^= 1;
6810 reverse_code = 0;
6811 }
6812
6813 op0 = XEXP (x, 0), op1 = XEXP (x, 1);
6814 if (earliest)
6815 *earliest = prev;
6816 }
6817 }
6818
6819 /* If constant is first, put it last. */
6820 if (CONSTANT_P (op0))
6821 code = swap_condition (code), tem = op0, op0 = op1, op1 = tem;
6822
6823 /* If OP0 is the result of a comparison, we weren't able to find what
6824 was really being compared, so fail. */
6825 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
6826 return 0;
6827
6828 /* Canonicalize any ordered comparison with integers involving equality
6829 if we can do computations in the relevant mode and we do not
6830 overflow. */
6831
6832 if (GET_CODE (op1) == CONST_INT
6833 && GET_MODE (op0) != VOIDmode
6834 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT)
6835 {
6836 HOST_WIDE_INT const_val = INTVAL (op1);
6837 unsigned HOST_WIDE_INT uconst_val = const_val;
6838 unsigned HOST_WIDE_INT max_val
6839 = (unsigned HOST_WIDE_INT) GET_MODE_MASK (GET_MODE (op0));
6840
6841 switch (code)
6842 {
6843 case LE:
6844 if (const_val != max_val >> 1)
6845 code = LT, op1 = GEN_INT (const_val + 1);
6846 break;
6847
6848 case GE:
6849 if (const_val
6850 != (((HOST_WIDE_INT) 1
6851 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
6852 code = GT, op1 = GEN_INT (const_val - 1);
6853 break;
6854
6855 case LEU:
6856 if (uconst_val != max_val)
6857 code = LTU, op1 = GEN_INT (uconst_val + 1);
6858 break;
6859
6860 case GEU:
6861 if (uconst_val != 0)
6862 code = GTU, op1 = GEN_INT (uconst_val - 1);
6863 break;
6864 }
6865 }
6866
6867 /* If this was floating-point and we reversed anything other than an
6868 EQ or NE, return zero. */
6869 if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT
6870 && did_reverse_condition && code != NE && code != EQ
6871 && ! flag_fast_math
6872 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT)
6873 return 0;
6874
6875 #ifdef HAVE_cc0
6876 /* Never return CC0; return zero instead. */
6877 if (op0 == cc0_rtx)
6878 return 0;
6879 #endif
6880
6881 return gen_rtx (code, VOIDmode, op0, op1);
6882 }
6883
6884 /* Similar to above routine, except that we also put an invariant last
6885 unless both operands are invariants. */
6886
6887 rtx
6888 get_condition_for_loop (x)
6889 rtx x;
6890 {
6891 rtx comparison = get_condition (x, NULL_PTR);
6892
6893 if (comparison == 0
6894 || ! invariant_p (XEXP (comparison, 0))
6895 || invariant_p (XEXP (comparison, 1)))
6896 return comparison;
6897
6898 return gen_rtx (swap_condition (GET_CODE (comparison)), VOIDmode,
6899 XEXP (comparison, 1), XEXP (comparison, 0));
6900 }